Editor’s Note from Mike: BRXND NYC is 50 days away and as promised, we’re continuing our interview series with leaders who are taking the stage at the event.

In today’s edition, Noah chats with Pip Bingemann, who purpose-built a model designed specifically to inspire, not give you the answers. To see Pip showcase this live alongside our incredible set of speakers building on the frontier of AI and marketing, grab a ticket to join us in the Times Center on Nov. 5 here.

Buy Tickets Here

On Divergence Models

Back in May, I talked with Pip Bingemann, the co-founder of Springboards. He showed me Flint, a small fine-tuned model they built because the frontier models all tend to give us the same stuff.

The canonical example he likes to cite here is to go ask ChatGPT or Claude for a random number between one and ten, and you get seven. But when you ask for a car, you get a Camry or a Civic; or when you ask for a running shoe tagline, you get “Run your way.” Despite being developed by different companies, they all come back with the same answers.

Springboards had already seen this in its creativity benchmark, where 16 models answered briefs for 100 brands and more than 600 people judged the work.

Pip calls this convergence. I don’t think this convergence is purely bad: sometimes you need an average answer, like when you’re doing synthetic research. But anyone who has tried to use AI for creative work can attest to the fact that it can be difficult to get where you need (especially if you need to be funny).

That is where Flint comes in. It is not supposed to do everything. Springboards calls it a “divergence model.” NoveltyBench basically asks the same open-ended question 10 times and counts how many meaningfully different answers a model gives. Flint scored 7.47, compared with 2.54 for GPT-5.4 and 1.83 for Claude 4.6 Sonnet.

Since we talked, Kieran Browne has written a technical explanation of Flint, and Will Douglas Heaven has covered the model for MIT Technology Review. This is lightly edited for readability, but it is basically the conversation we had a few months ago. I hope you enjoy it.

Noah Brier: I’m very excited to talk about Flint. I think we are converging on a similar set of ideas. The thing I have been thinking about is that I pretty firmly believe the models are getting worse at writing, not better. My feeling is that RLHF fundamentally stands in opposition to good writing.

Pip Bingemann: I think reinforcement learning is 100% part of it. But I think there is a much bigger issue at play.

Flint is not a foundational model. It is a fine-tuned model. It is a very small fine-tuned model. There are problems with it.

The thing we see is that the models are highly predictable because they converge into the same space. The answers are almost always the same. If you ask for a car model, you will probably get Toyota Camry, Toyota Corolla, Honda Civic, Ford Mustang, maybe Ford F-150. Every time, they land in the same space. Flint gives you something else.

Or ask for a tagline for New Balance running shoes. It could be anything. What you tend to get is “Run your way” or “Run your world.” Flint gives you something different every time.

For me, this is the scary thing. If I copy the same prompt into a new session again and again, it is the same over and over. And I am not picking on Claude or GPT. If I put Gemini in there, it is the exact same thing.

Noah: So what is Flint, exactly?

Pip: Flint is a model. It is a tiny model, a 30-billion-parameter model, but it breaks the distribution of predictability you see in every other model.

It is not good for everything. But when someone asks an open-ended question, they deserve an open-ended answer. At the moment, you are not getting it. People ask, “Can’t you just dial up temperature?” You cannot.

The way we use it in Springboards is that we do not believe Flint is a good model for everything, but it is good for some things. We give people the option to switch between models. If you want to put on a divergence model, you can. Or you can leave it on auto and we route to a model based on the request.

Noah: Why do you think this is happening?

Pip: There is one paper you need to understand: Artificial Hivemind. The basic finding is that regardless of how these models are trained, they have all converged into the same space.

One of the tests was: give me a metaphor for time. Of the 25 different models they tested, nearly every one said some version of “time is a river.” The models are highly predictable and repeatable, but they have also converged into the same space.

If you ask for a motto for a social media page, you can get word-for-word the same answer from very different models trained in different ways. I think it is a very scary thing.

The reason I think this is happening is a couple of things. We have already run out of data. There is more AI-generated content in the world every day, and it is going back into the world. At the same time, the quality of our data is degrading. The models are eating themselves because they are training on the open web, which is increasingly AI-generated.

Synthetic data is not bad. It depends what it is for. In worlds where there is a correct answer, that is fine. But in worlds where it is open-ended, it is not okay.

There is another paper about model collapse when models are trained on recursively generated data. There are two sentences from that work that I love. One is that probable events poison reality, because the probable and average get overestimated. The average doubles down.

The other is: the tails of the distribution die first.

That is what I think is happening. As these models get bigger and are trained on the same data, the tails shrink. They get pushed into the exact same space over time.

Noah: I buy that as the underlying explanation. My challenge is maybe a layer above it, which is that I am amazed how hard it is not to generate slop. It is so insidious.

That said, I am not sure I think convergence is an overall bad thing.

Pip: It is not. It depends on the use case. That is why I say Flint is not good for every use case. We will never be 100 percent Flint because it does not make sense. Sometimes you want the average.

Noah: Yes. In a lot of cases, getting to average is totally fine. For some marketing cases, it is totally fine. It is actually why AI can be pretty good for synthetic research. You exactly want the average and you do not need to push against it. But on the creative side, that is why there is work to do.

Pip: Exactly. I liked that sentence: sometimes getting to average is fine. The quicker people realize that, the better. It is an improvement in a lot of cases. For a lot of people it is, and for a lot of people it is not.

Noah: What are you doing with Flint now? Is it just inside Springboards?

Pip: Right now, yes. The more important thing is probably what is wrong with it.

It is a 30-billion parameter model, which is tiny. With a smaller model, there are two big problems. One is that its world knowledge is not as good. The other, bigger problem is that it struggles with complicated prompts. It has a big context window, around 250,000 tokens, but its coherence of instructions collapses pretty quickly.

At the moment it is released as an alpha model inside Springboards. We route to it occasionally and monitor how much we route to it. We let people switch to it if they want to. At the same time, we are scaling the model up and building it as an API.

Noah: Part of the reason I ask is because I would love to try it when you have the API. We have a bunch of those shape problems we need to solve, where we are specifically looking for divergence. We are playing with all these techniques for getting to not boring.

Pip: One of the tests we use for Flint is: tell me a dad joke.

If you ask Claude or GPT to tell you a dad joke, you get the same joke every time. Claude is always atoms. GPT is always skeletons because they do not have the guts. Sometimes it is a scarecrow. You cannot be funny if you only have one joke.

With Flint, we actually give you different jokes every time. There is a repetition problem and a convergence problem, and Flint kind of looks after both.

The same thing happens with prompts like “pick an emoji at random” or “pick a number between one and a hundred.” It is always 73 or 42 or 47. With Flint, we actually give you randomness.

Our point with this is that I do not care about quality in the first step. I really do not. It is not up to me to decide what is good and what is bad. That is your job. I want to help you expand your horizons and spark an idea in your head. Then it is up to you to work out what is good and how to overcome the rest.

You just cannot even do that with these other models. That is the scary thing.

Noah: How has it been building a model? Painful?

Pip: It has been really good. We have been working on it for just under two years.

The cost to build the model was cheap. The compute was cheap. The expensive part was all the experiments to work out how to get this thing to drive divergence without degrading the base performance of the model.

That is the real breakthrough with Flint: how do you get this thing to drive divergence without degrading the base performance of the model? We give it knowledge benchmarks as well, like MMLU and science tests, and we do not get worse on those tests. We actually drive variance without degrading knowledge.

That was the hard part, and that is the fun part. It is a never-ending job. It is my favorite meeting of the week because you go: here are papers, let’s try it. It does not work. Here are the problems with our model. Let’s scale them up and keep going.

Noah: That is awesome. I am excited to try it when you put it on the API. I think these are very interesting questions. Even your pushback that Flint cannot do everything feels right to me. I do not need one thing that can do everything. I need all these pieces. I need all the ingredients.

The problem for you is that people have to know enough to know this is a thing they need to solve for. With randomness and divergence, the vast majority of people have not approached the edge of any of these models yet.

Pip: This is the scary thing, particularly with advertising. I sent a message to a friend who is a pitch consultant and asked, “As you help clients go through the pitch process, are you noticing responses becoming more similar?”

He said, “More similar? Almost a joke. In the last 12 months, the net effect is the supply chain gets devalued. It is not good for anyone. It is hard to argue for nuanced premium and margin when everyone’s giving you the same answers.”

That is the problem no one realizes. The advertising agency world, in particular, is walking into it blind.

Noah: That one goes deep.

I have had three friends over the last few months who work at different places where their agencies have pivoted to be “AI-first.” And they are like, what does this even mean?

To me, I am just like: if you were any good at marketing, would you not recognize that counter-positioning is a pretty critical thing? If everyone is saying the same thing, there is probably a real advantage in being counter-positioned against that thing instead of saying you are something you very clearly are not.

I feel for them, but these models are sucking so much oxygen out of every room that you are left trying to fill it in. But the idea that you are going to fill it in by saying the exact same thing as everybody else seems like a problem.

Is there anything more disqualifying for an agency than not being able to put together strategic positioning for yourself?

Pip: It is insane. Especially because their entire history is that their job has been to sell people. That is what they do. They sell the talent of their people.

Then to say, “You do not need people, you just need AI,” is insane.

Noah: Exactly.

Pip Bingemann is the co-founder and CEO of Springboards. You can read more about Flint here. To see him live on stage at BRXND NYC 2026, click here to purchase tickets.

As always, thanks for reading.

BRXND NYC is 52 days away and as we get closer to the big day, I’ll be turning over more of the newsletter mic to our outstanding speakers and sponsors who are at the heart of how AI is fundamentally reshaping the way consumers build preference and purchase brands.

Today, we have a conversation with Justin Inman, Founder & CEO of Emberos, the operating system for AI brand management and a sponsor of this year’s show.

Justin will be on stage November 5 at the Times Center along with leaders at Amazon, Rocket Money, Bark, Microsoft and a whole bunch more being announced soon. Click the button below to grab your ticket or reach out to me at mike@brxnd.ai with any questions.

Buy Tickets Here

On AI Brand Management

“Knowing what ChatGPT says about your brand is useful. Knowing why it says it is more valuable. Knowing what is likely to change next, what you can do about it and whether that action worked is where this becomes an enterprise operating function.”

This quip is essentially Tom and I’s AI Search Maturity Index in a pithy 40-word nutshell. It’s also the final line in the press release announcing that Emberos, the operating system for AI brand management, raised a $5.5M seed round earlier this month.

Emberos is a newer player on the AI search scene but one I’ve followed closely since inception and marveled at the ambition around. Emberos operates a multi-agent platform that reads how AI represents a brand across the major LLMs, predicts outcomes, orchestrates fixes automatically, and proves lift through the Brand Knowledge Graph and Tiered AI Visibility Index. All in a company that is only a year and change old.

Thus, I’m ecstatic that Justin Inman, Founder CEO of Emberos, is taking the stage at BRXND to talk about how online discoverability has shifted in an AI/LLM driven search world and what brands can do to manage it proactively.

A former enterprise growth leader at Google, Justin was pivotal in helping major brands such as Amazon, Sony, Coca-Cola, and L’Oréal navigate marketing’s last major upheaval: digital transformation. He also serves on the IAB’s AI Board, where he shaped Measuring Visibility in the AI Era, the IAB’s industry playbook for measuring brand and publisher visibility across AI-powered discovery platforms.

Before he takes the stage at BRXND, Justin and I caught up to talk shop a bit on the current state of how AI perceives brands, the concept of a “brand “knowledge graph,” and the rising importance of creators in shaping a brand’s destiny with LLMs.

A lightly edited version of our conversation follows below:

Mike: You position Emberos as “brand management for the agentic era.”

I suppose my light devil’s advocate comment is that your brand should always have been consistent across channels, and this is just good hygiene with a new name on it.

How are the stakes and mechanisms of brand management different now than they were in the proverbial before LLM times?

Justin: Brand management has been around forever and it’s going to continue to be around. It’s just going to look a little different.

For the first time ever, marketers and brand managers have to think about a couple of new things. The first is they’re marketing to machines and bots. That hasn’t historically existed and it’s a different audience they’re going after.

The second piece is that the point of discovery is incredibly fragmented now. If you’re looking at this new world of AI chatbots becoming the front door of the internet for the first time ever, brand managers have to rethink that marketing funnel across awareness, consideration, intent, purchase. That’s all happening through a single layer now. The playbook of pushing people down the funnel looks very different in this world than what they’ve been trained up or educated on.

Brand managers historically marketed to different channels intentionally, because they knew different audiences lived on those channels. There might be different messaging on social versus your site versus PR, because that was intentional. The problem is that breaks in this world, because the LLMs favor coherence. The second you start to have different messaging, it gets confused, and that’s where we get into the hallucinations and all the other things that come with it.

In five years we probably won’t be calling it brand management for the agentic era, or AI brand management. It’s going to be just brand management, because it’s going to naturally encompass this.

Mike: So what does AI brand management look like at the CMO level?

Justin: CMOs are starting to realize that this is a brand coherence issue, and it has to touch not just your typical marketing teams, or the digital team, or the search team, or even the tech team. They all need to be orchestrating together.

They’re feeling the pain. They see their organic traffic getting crushed. They’re going into these LLMs themselves and seeing the positioning and the narrative across their brand and realizing it’s inaccurate.

Mike: AI brand management is a bit different than some of the more industry standard terms like generative engine optimization……but I get the sense that is intentional.

How do you counter-position Emberos against some of the other platforms enterprises are buying that have been in the market a bit longer than you have?

Justin: A lot of these companies are basically saying, here’s a box, play in it. We know, because we work across a bunch of different verticals, that what works within gaming is very different from what big beverage wants or banking wants. It’s a different universe, different relationships. So it’s a disservice to any customer using an off-the-shelf tool, they’re cracking the window, barely looking through, and not seeing the full picture.

A lot of the big platforms today are also looking at this as a content generation solution, you have to create more and more content to compete into these LLMs. They’re looking at it from a concept of push-to-publish, which is AI-generated content, which is AI slop in our opinion.

These are short-term visibility hacks with potentially long-term implications for your traditional search business. It’s a content treadmill that you will never be able to get off of. We don’ tdo that.

The Brand Knowledge Graph allows us to be way more focused on what you need to do. We look at the most highly authoritative cited sources and make optimizations against those, versus just more and more slop and content.

I’d also say a lot of platforms give you a laundry list of optimizations and never tie it back to any type of meaningful lift. At Emberos, we give you the biggest rocks to move with a predicted lift attached to it. We track those changes in our platform and then we show you the actualized lift. We’re the first closed-loop platform in that respect to hit the market.

Mike: Break down the Brand Knowledge Graph further for me. What are the core inputs, and how does that change how a model mediates brand preference?

Justin: We can map over brands, products, creative assets, influencers, creators — you name it, it doesn’t matter. Once we map that over to our systems, we build up a node structure, and that node structure is dependent on the prompting layer. We essentially develop a city map of how the internet looks at that specific entity we’re tracking.

So we can see all the biggest influencing factors online, the publishers, the creators, the site itself, and how those all correlate to the entity we’re tracking. The cool thing is we can do it by the ecosystem of that brand. But we can also look at it at the vertical level, or from a competitive set.

Then we track all those relationships over time, which gives us a baseline, which unlocks the predictive capability. So we can say, if you make this change to this site, or get on this publisher, or work with this creator — because of that baseline, we predict that X outcome will happen.

I’d also say the tools today have not historically given a lot of insight for the why. They’re just giving you the how. The more advanced marketers we talk to have been working with a tool for X amount of time and there’s still a gap in fully understanding the reasons why competitors are gaining visibility.

Mike: A lot of your early clients are in entertainment and media where more than most categories, brand management is paramount.

What’s the canary in the coal mine for how media and entertainment businesses navigate the transition to AI search for marketers in other industries?

Justin: Entertainment’s an interesting one, and I learned this in my time at Google. It doesn’t matter what industry you’re in, every marketer looks at entertainment to find out what they’re doing. It’s partially because they have their finger on the pulse of culture better than any other industry.

They’re going to LLMs and seeing, for streamers, attribution in terms of where to stream — you’re Hulu and you’re seeing “stream this on Paramount+,” which is a massive issue. If you’re a studio, LLMs misgenre about 20% of the time.

The LLMs get confused about what these titles actually are about. If you have a title with the word “dead” in it, it automatically thinks it’s a thriller. But it might not be. So you have titles getting the wrong genre, wrong actors, wrong directors, or even wrong dates because the movie industry does a soft release and then the box office opening weekend, and if the model picks up the soft release, it’s wrong.

If moviegoers are asking what they should see this weekend, and the genre is labeled as a violent thriller but it’s actually a dark comedy, it alienates a whole portion of your population that could go see that movie. And it sticks. Six months later someone says, I liked this movie, give me 10 others. You need to be in the right genre and associated with the right comp set.

Same thing with gaming. If there’s Grand Theft Auto 1, 2, 3, 4, it’s very hard for the models to tell the differences between each one. Gamers are asking specifics about a specific title, and a lot of times it’s incorrect character associations, incorrect where to play, where to buy.

Mike: Talk to me a bit more about governance because I believe that will ultimately be the most important battleground for AI search, especially in the enterprise. Why is monitoring truth and compliance across AI such a hard problem and how does Emberos help brands here?

Justin: Every AI answer is generated on the spot, and it comes out different depending on the model, the user, and the day. There’s nothing fixed to review and nobody inside a brand currently is set up to own this governance. Legal owns claims, comms owns narrative, SEO owns content, and AI blends all three into one sentence that’s in nobody’s job description.

We created Echo for this reason. It validates what the models are saying before anyone acts on it, traces where the information came from, flags where they contradict each other, and keeps a record of who approved what. It’s a peek inside the black box of LLMs before you have an issue that you are chasing to fix.

That matters now because brands are moving from watching what AI says to acting on it. Once you’re making changes at that scale, you need to know the information behind them was accurate, and you need to be able to show what changed and why.

Mike: In a similar vein, you’ve been running a live public test of your predictive claims on box office data. Walk me through this and how this work might map to the larger zeitgeist of agentic commerce.

Justin: Back in Q4 we said, can we test our predictive capabilities at a macro level? We looked at a space that’s rapidly changing and allowed us to make a prediction and then relearn and retrain. That space was the box office.

So we have made it a practice. On Fridays we do a box office prediction, on Monday we see what the normalized numbers are, and then we retrain our models. What we’re looking at is how these titles show up in LLMs, and we’re seeing a strong correlation to box office revenue — the types of prompts people are asking, what we call Share-of-Prompt, and how that correlates to real-world business outcomes.

We see signal lead times almost 12 days before traditional Hollywood trackers. And in some cases our box office prediction is two to three times more accurate than traditional Hollywood trackers.

We’re not a box office prediction company, but we’re going to replicate that model across our priority verticals. The ultimate goal is to get enough reps under our belt to say, a 1% Share-of-Prompt increase equals X amount of box office. It gets really compelling when you think about SKU-level optimization and the agentic shopping world.

What we’ve also seen is companies that historically relied on Google search feeds to update their forecasting models — it’s broken, because AI Overviews and AI Mode are no longer a clean feed. We have companies scoping us to see if we can fill that gap.

Mike: So zooming out, where do you see your suite of tools most producing genuinely counterintuitive insight for marketing leaders to take action on?

Justin: What we’re doing around creators, those examples are an aha moment for marketers. We just worked with a running shoe company and said, here are the creators you’re currently working with. We mapped those over and said, these creators are actually getting cited in AI, but they’re not getting cited for your specific company. So that’s an issue. You need to restructure the creator content — and here’s the list of things you need the creator to say that will help get them cited more.

We’re rolling out something called a Creator Graph right now. More and more, the void is going to be filled with creators, because there’s not going to be enough content as publishers start blocking the labs to figure out their true citation influence before they renegotiate. But the issue with the creator stuff is we’re seeing creators who aren’t even creators influence models.

It’s a mom in a car with her daughter saying, I just gave my kid these vitamins, they’re great. And the video has all of 100 views!

Mike: What about the creators themselves and their representatives? How much are they starting to care about how their creator brand is mediated by AI?

Justin: Often, talent is driving influence over AI answers and the talent doesn’t even know it. Creators are always looking at shares, potential reach, maybe affiliate sales. I think the next new metric they’re going to be looking at, probably in the next 12 months, is citation. How much they’re actually driving the influence over the LLM output.

The bigger agencies are starting to wake up that the agents themselves are doing all the research for their clients. So if their clients aren’t represented correctly in these LLMs — where everyone is researching projects and potential brand partnerships — that’s a big issue.

Mike: Close this out for me in style. With a fresh round of funding in the bank, talk to me about what’s next for Emberos, particularly at the nexus of creators and AI search.

Justin: We’re constantly expanding the Emberos Brand Knowledge Graph, adding new models, mapping new categories, and making our predictions sharper. But what excites me most is our work with creators. It’s genuinely category-defining.

We can trace an idea from the creator who sparked it, through the conversations that gave it momentum, all the way to the AI-generated answer a customer ultimately reads. Then we help brands understand that journey, participate in it authentically, and shape it for maximum impact.

We built Emberos as an operating system for a market that refuses to stand still. As the relationship between creators, brands, and AI search evolves, Emberos evolves with it. That’s the kind of partner we want to be: not another platform brands have to keep up with, but the reason they stay ahead.

Justin Inman is the Founder and CEO of Emberos, the operating system for AI brand management. He’ll be on stage at BRXND NYC 2026 on November 5th at the Times Center.

To suggest additional builders that I should feature in this series, please get in touch at mike@brxnd.ai

As always, thanks for reading.

— Mike

Purchase Early Bird Ticket

Muse & The Agentic Commerce Conundrum

Agentic commerce is suddenly very real folks.

Two weeks after Instinct initially took the tech world by storm, the hills of the Twittersphere are alive with the sound of people executing complex purchases end to end with a few simple texts in iMessage. Almost overnight, we’re seeing the inklings of a true agentic shopping experience that will really feel superior to many shoppers and not just on commodity, emotionless purchases.

Now, as he so often does, Zuck has come in off the top rope as a fast follower with a personal agent called Muse that will quickly bring this technology to the masses. Early reviews suggest that Muse is very, very good across both life admin and agentic commerce use cases. A month ago I asked when consumer AI would have its “Claude Code moment.” This is it. Yet again, the road to financial run is paved with people who bet against Meta.

With each passing day, it feels increasingly likely that we’ll work primarily within harnesses in the next 12-24 months. Now, if Meta can shape the future in their image and likeness, we’ll run our personal lives and shop in them as well.

There are a plethora of fascinating second order questions to spill ink over here but top of mind for me is how brands of all stripes respond to this sudden influx of consumer agents. We’ve already seen the first flashes of tension with Resy deactivating user accounts after Instinct made 200+ requests in an hour. This is a perfectly reasonable response and microcosmic of a much larger point of tension.

Simply put, the best agentic commerce experience for users is going to often be at odds with optimizing yield for merchants. Tools like Instinct, Muse and Grok are going to mercilessly hunt for promo codes and off-menu discounts and will consistently try to “beat the system” in unforeseeable ways on behalf of their users. But beyond that, what really is going to make many of the largest incumbent platforms on the internet viscerally uncomfortable is that they no longer control the shape of a buyer’s purchase journey.

Agentic commerce breaks many of the meticulously optimized purchase funnels (often at razor thin margins) that consumer businesses have spent the last two decades building online. It’s also incredibly unclear if or how entities like Muse and Instinct will be influenced by traditional sponsored product ads as they decide what to purchase. This is going to cause a major conundrum for merchants, especially because from a profit perspective, most large retailers are now advertising businesses masquerading as purveyors of goods.

Brands are understandably skeptical of the version of the future in which they get disintermediated from their customers. There’s not much nobility in being a wholesaler and logistics arm…and you never beat Amazon at that game anyway. I’ve long posited that many have effectively been playing a version of narrative arbitrage where they are enthusiastic about agentic commerce in public to appease the street but behind closed doors, they are happy to see early pilots struggle to gain traction. Walmart enthusiastically backed OpenAI’s early embedded checkout experiments but was pretty quick to publicly share how much lower the average yield was off their own platforms. That “agentic commerce converted 3X worse” number sure spread in the press like wildfire.

That said, I’m far less convinced then Eric Seufert that open agentic commerce outside of retailer owned ecosystems is entirely a “mirage”. Blocking off your marketplace writ large to an agent eager to quickly buy something from you (and likely operating on behalf of a person who spends a lot more money annually than the median shopper) is cutting off your nose to spite your face.

It’s far better to be early to a platform shift that is happening gradually than late to one when it happens suddenly.

The Three Miracles of Agentic Commerce

Historically, the macro knock on agentic commerce has been that it further solves buying but kills shopping in the process.

Last month, I put forth a framework that inspired lively discussion on LinkedIn around the “three miracles” that need to happen for agentic commerce to ultimately be better at helping users discover what to buy than the status quo experience in online commerce today.

1) Merchants must expose enough structured data and tell a consistent brand story at both at the catalog level and across the web for agents to understand how individual products match to highly semantic and fuzzy intent.

2) When items are sold at multiple retailers (or sometimes when choosing what to recommend in first place), agents must consider which merchant provides the best overall value as a function of price, loyalty status, rewards, delivery window, and more.

3) Agents must possess enough “media literacy” to understand which data sources are reliable arbiters of product quality. In other words, they must discern pay to play slop masquerading as organic product reviews from publisher or creator content where an objective reviewer actually tests a product.

While #2 and #3 are largely in the domain of the companies building agents, point #1 is directly on brands and by far the most impactful lever marketers have to control their destiny in agentic commerce, at least until we can all buy ads to influence agents at the point of inference.

While there’s much handwringing from the tech Twitterati about if agents should be broadly anthropomorphized, I strongly believe that agents will effectively be more extreme versions of human buyers when it comes to gravitating towards brands that are story rich and consistent with their messaging across all key touchpoints. As Noah wrote back in December:

”When you RLHF a model to be a “helpful assistant,” you’re essentially training it to care about context, explanation, and story (emphasis mine)—exactly the things we thought the robot would skip past.”

What I’m really trying to say here is that however big a seat your company thinks the marketing team should have at the agentic commerce table, make it bigger.

Shopfiy’s Field CTO Rohit Nathany put this cleanly: “I highly recommend my marketing and growth friends start treating catalog and product data as a marketing tactic, not just an admin task. Your catalog should be complete, structured, and story-rich.”

On that note, I’d love to jam with any marketers who are currently working on this problem— drop me a line at mike@brxnd.ai.

The Brand Brain & Turning Context into Competitive Advantage

As more work moves into AI tools, a company’s advantage comes from how effectively those tools can use its own knowledge and expertise. Said another way, context is king and a “brand brain” is how companies turn that context into an enduring competitive moat.

With assistance from Youtube maestro Sam Knight, Noah has been on fire lately on the video and podcast circuit and we’ll be bringing much more of Noah and Sam’s work to you here.

In the video below, Noah unpacks what a Brand Brain looks like in practice: connecting a company’s data, context, expertise and tools so AI can work with the knowledge that actually makes the organization distinctive.

Geno Smith vs. Armageddon

Football is back….which means I once again have to remind myself that suffering through another New York Jets season is a choice I freely make. Nobody is forcing me to do this! I can pull the plug at anytime if this isn’t even achieving the goals I want it to.

Anyway, if like me you’re mildly distressed that an Anthropic staffer seems to genuinely believe that there is a 10% chance AI will cause mass extinction and he’s powerless to stop it, here’s a quick post to put your mind at ease.

I would at least like to see one 34-13 wild card round loss to the Baltimore Ravens on a miserable sleet-soaked January afternoon at MetLife before AI brings upon my imminent demise. Is that really too much to ask?

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

Purchase Early Bird Ticket

Since Noah published “On AI and Writing” two weeks ago, the discourse around AI as a “legitimate” medium for communicating original ideas has only intensified. Tim O’Reilly wrote a contrarian defense of writing with AI arguing it is “a field of possibility from which humans can and will summon up and communicate insights, truth, beauty.” Stanley Druckenmiller proudly published a Wall Street Journal op-ed with AI and was resoundingly backed by the paper’s editors. On the flipside, Derek Thompson shipped a phenomenal interview with Pangram founder Max Spero pleading that “we can’t let AI writing take over the internet.”

Suffice to say, for folks like me who attach a significant amount of their professional identity to being “writers,” it’s been an especially intense two weeks where it’s become ultra clear that the genie isn’t going back in the bottle.

Like Noah, I started my career as a journalist and maintained an active second career as a ghostwriter, blogger and newsletter author throughout my decade or so working in technology companies. No matter what I was doing in my day job at the time, any writing project was always the “hardest” work I was doing. Staring down the barrel of the blank page burned more intellectual calories than anything else I was working on.

This dynamic has only gotten more intense in the age of AI. LLMs are much better than I am at “low-caloric” writing tasks, leaving only the arduous work of thinking in their wake. Therein lies the agony and the ecstasy of this moment for anyone who writes words online for a living. The job is harder than ever but the work that remains is the most gratifying aspect of why we do this.

Building on Noah’s piece, here are a few thoughts I have for writers trying to future proof their craft and understand their raison d’etre as AI eats the world.

1. While I agree with Noah that the more powerful models are getting worse at writing in many respects, I believe they have crossed the threshold of being better at the “view from nowhere” style of summation analysis than even great human writers. For an omniscient take on most subjects, AI-generated text does the job at hand. You’ll probably even learn to love some AI-generated work.
   
   So what remains scarce when the majority of what has been written for the web is done well by LLMs? Insights derived from true shoe-leather reporting, analysis drawn from learned and felt experience and things that make people laugh. Writing that doesn’t check one of these boxes will be owned by a machine.

   
   

2. The original promise of the internet is that it was supposed to free writing from the rigid form-factors of newsprint and allow unbounded creative expression. In practice, so much writing for the last two decades has been self-censored to fit in the SEO box. We are increasingly being freed from this.
   
   However, the tough thing for writers adjusting to this new world is that before LLMs, there was an enormous amount of demand for “view from nowhere” takes optimized for SEO. The market made us into one thing and is now aggressively demanding another.

   
   

3. In a similar vein, Sam Parr has a really good “barbell” theory of internet writing that the best posts are either <500 words or >2K word long essays. His advice is to avoid the messy middle length that defines much internet writing. This is more true now than ever before.
   
   It is also really hard! The ~1,000 word messy middle is the easiest form factor to make an argument in and generally where the narrative wind blows me. This is largely due to the fact that for two decades, 800-1,200ish words was the standard length of piece commissioned by editors for op-eds and many manners of freelance writing.
   
   I am deeply struggling to reframe my craft around 400-500 word quips, supplemented by the occasional piece d’resistance type essay. Note the length of this piece! But I believe this to be a near existential thing to get right.

   
   

4. I agree with Noah’s take that the best writing is fundamentally anti-consensus. But this presents a conundrum for writers who have to ship content on a consistent cadence. The reality is that you just aren’t going to have that many good, contrarian takes! So how else can you consistently write pieces that can’t be replicated by AI?
   
   Many of the best pieces of internet writing are something that an industry, subculture or at times an entire society is deeply feeling but lacks the precise eloquence or courage to coherently say out loud. I believe this “finger on the pulse” style of writing will be safely LLM-proof for a long time and will become even more valuable as emotionless AI-generated copy eats the web. For now, models simply cannot “feel” vibes in real time at nearly the level people can.

   
   

5. Lots of the best internet writing is going to look increasingly rough around the edges. I believe we’ll see a style emerge that looks less like traditional essays and more like “jagged collections of bangers” thrown together without the usual transition paragraphs.
   
   This will be a response to both reader demand for brevity and a counterpositioning against the tendency of LLMs to create flowy, overly cohesive narratives. It will be weird for a while but ultimately liberating for readers and writers alike.
   ‌

6. For my money, Matt Levine is the consummate example of the most AI-proof writer on the web. If you want to have a good laugh, try prompting a model to write an essay on a topic you know well “in the style of Matt L’s Money stuff.” AI is horrifically bad at capturing Levine’s je ne sais quoi of wit backed by a couple decades in the boiler room.
   
   Honorable mention here goes to Sam Kriss as a sheer master of prose that operates on a level that simply can not be emulated.

   
   ‌

7. To steal from Picasso, LLMs are going to be better than you at “learning the rules like a pro.” They will not be better at breaking them like an artist.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

But first, BRXND NYC is 65 days away on 11/5, and tickets are currently $749. We look forward to seeing you at the Times Center for a great show.

GET YOUR EARLY BIRD TICKETS

The Three Vectors of the AI Race

Increasingly, I believe there are three primary vectors in the AI race:

1. Raw IQ: Who has the smartest model?

2. IQ + Cost: Which model is the smartest per token dollar?

3. IQ + Cost + Speed: Which model is the smartest per token dollar at the highest tokens-per-second (toks)?

While raw IQ is the frontier race everyone pays attention to, I think IQ + cost is the most interesting thing to watch in AI right now.

But before we get there, let’s talk about raw IQ. The consensus (and my personal POV) puts Fable on top, with Sol behind, then a fairly big gap to number three. Broadly, though, this class of model should be reserved for your toughest problems where the costs don’t matter that much. If the AI is solving something truly valuable, then the $10/50 (input/output cost per million tokens) you spend on Fable shouldn’t matter that much. The problem, of course, is that the way most tokens are getting consumed today are in coding harnesses like Claude Code and Codex, and when you’re doing that kind of work, a lot of people just leave the smart model on and let it cook. Luckily, these harnesses, particularly Claude Code, are getting smarter about offloading easier tasks to cheaper models in the form of subagents. A lot of people dream of routers to solve this, but I think for the most part we’ll be good with this smart orchestrator, dumb-agent architecture.

More interesting, though, is that even today I’d argue the vast majority of work can now be done by much cheaper models (Luna, particularly in Max mode, is an amazing workhorse that’s worth spending more time with). Which brings us to bucket two: IQ + cost. Lots of people are talking about this as the Pareto Frontier, and I think it’s the most interesting space to watch right now. Segment co-founder Calvin French-Owen had a great post titled “Small Models Have Arrived” where he discussed the value of this class of model:

I think demand for “frontier-level” models is going to keep compounding. Especially for fields that require novel breakthroughs or discovery (engineering, hard science, model training).

But I also think the demand for “fast/cheap/good-enough” models is just about to take off.

Think of the people you interact with on a daily basis: coworkers, vendors, and customers. Nine times out of ten, you want someone who is super responsive, and just handles things for you. Most of the “human tokens” at companies today are spent this way — hiring skews heavily toward the fast/cheap/good-enough archetype.

This is the most competitive race in AI by far. Two labs are competing for the true frontier, but at least eight serious entrants exist in this space thanks to Open Source. These are the workhorse models that power the real things happening behind the scenes in agentic systems. At Alephic, we catch every Slack message from our company chat and classify it with Google Flash-Lite, the former leader of this model category. In just the last two weeks, GLM 5.3 Flash and Qwen 3.8 27b have emerged as serious competitors here. When OpenAI dropped the price of Luna by 80% (!!!!) last month, it became the de facto leader here. Back in March, I described this as the $1 sweet spot, where the blended average cost per million tokens was under one dollar.

To Calvin’s point, I also believe this space is about to explode. Most inference shouldn’t be happening on the absolute frontier. Using Fable to reply yes to a scheduling email is, to understate things, not the best use of its capabilities—let alone having it do basic classification tasks like “does this Slack message require further action?” The fact that so few people talk about this mostly indicates that most folks haven’t actually got agents running in production, so their primary experience with these models is through the harnesses, where it’s almost always easiest to leave the smartest model on by default. But as soon as you have to ship a system and you care about cost per task (and speed), you naturally spend time finding the best model at the right intelligence/price point.

Finally, the last vector is the least explored: IQ + cost + speed. In this category, I’m not talking about slightly faster, like going from 200 to 400 tokens/second (toks)—I’m talking about something like what Cerebras has with GPT-OSS 120b, which runs at 3,000 toks at a $0.35/$0.75 cost. This vector is the hardest to nail down because no one has a good intuition for what 3,000 tokens-per-second intelligence might be most useful for. I’ve played with some fun experiments with these models, like building an entire fake e-commerce store that generated pages and products as you searched, but the truly useful places I use these models are on problems where you want to insert AI into user-facing functions without them noticing. The best example is reranking: in our internal company CLI, you can search across calls, contacts, and companies, as well as Google Drive, and GPT-OSS on Cerebras reranks the results against the query as the final step before serving them back to the person/harness. Models at these speeds are still pretty dumb, but that will change, as will our imagination of how to use 3k toks intelligence.

As I said, I’m most interested in the cost + intelligence space, and I suspect the labs will push a lot of resources in that direction as AI maturity continues to grow. I also wouldn’t count out Google quite yet. While they don’t seem to have much momentum in the race for frontier intelligence, their bread and butter has always been squeezing performance-per-dollar out of their models, and I still find myself reaching for flash-lite even as new entrants emerge.

The final point is that as this cost + IQ space gets cheaper and smarter, we’ll see more stories like the one OpenRouter shared last week. When they discounted Luna and Terra they saw an explosion in usage. This is, of course, Jevons Paradox, and I’ve felt it in my own work. As I’ve used Luna more in Codex, I find myself doing many more tasks and a lot of long-running /goal work. If you think there’s a lot of inference now, just wait.

Since you’ve made it to the bottom, one more reminder that BRXND NYC is coming up on 11/5, and tickets are currently $749. Get yours early, we will sell out.

GET YOUR EARLY BIRD TICKETS

With 69 days (!) until our BRXND NYC conference, I’m continuing our series of Friday buildup interviews with founders, technologists and marketing leaders innovating in the problem space around how AI is fundamentally reshaping the way consumers build preference and purchase brands. If there’s someone you think I should feature, please drop me a line at mike@brxnd.ai.

More importantly, we’d love to see you at our November 5 show at the Times Center, Click the button below to grab your ticket!

Buy Tickets Here

On Agentic Commerce & Friction as Flame

“Life should be a feast. The algos have other plans.”

So begins Jess Graham’s Field Guide to the Age of Algorithms, a provocative almanac of phenomena at the intersection of AI, brands and culture. For my money, it’s the best resource for how marketers can navigate life after filterworld. In it, she coins terms like “the seamlessness paradox”, “discovery debt” and “beigeville” that now live rent-free in my vernacular.

Jess and I met serendipitously, two kindred spirits on a quest to push past technology into culture to consider what marketing looks like in a world where every decision is mediated by an AI model. Last month, I was doing the morning Substack scroll when a notification of a new follower and subscriber to BRXND stopped me in my tracks.

“I write about what happens when AI agents start making purchasing decisions for us and what is at stake for brands.”

Wait a minute, that’s what I do! Naturally, I reached out to Jess and we began to have wide-ranging jam sessions, one of which I bring to you all here.

By way of formal intro, Jess Graham runs Banquet, a practice for brands working out where their leverage goes in the age of brand disintermediation by platforms and agents. Prior to going independent, Jess held executive roles at Visa, Instagram and Meta and twice built consumer brands from zero as a CMO and GM. Jess spent nine years at Visa, running the consulting arm that worked inside banks and merchants making digital payments ubiquitous. At Instagram, she led the launch of Reels — working inside the machine that decides what people discover. At Meta, she wrote the first global data and privacy narrative after Cambridge Analytica. Not a boring career!

She now also writes The Feast on Substack for marketers, brand builders, and anyone who “suspects that efficiency isn’t the only thing worth optimizing for.”

Our friendly banter below is an amuse bouche of what she publishes regularly and has been lightly edited for clarity.

Mike: I want to start by having you answer your own existential question that you currently posit into the mean streets of LinkedIn as your profile headline.

What happens to brands when AI does the shopping?

Jess: When agents do the shopping, they are disintermediating people from brands. They are destroying relationships that consumers actually want to have with the things that they buy. They are removing all friction from the system by optimization. And that optimization has completely stripped out things like work, which creates dopamine, which creates attachment — just from a really fundamental biological human level.

Human centered design has never really been human centered design. It’s been lipstick on how do we make ourselves feel better about a commercial objective within an internet frame. If we were designing actually for humans, we would be introducing places where they have to think, where they have to do work, where they have to commit, where they have to raise their hands.

Having worked on a million different loyalty programs in my career, one of the things you learn really early on is that nothing is worthwhile unless I raise my hand for it. If you just hand me something, it’s not interesting and I don’t give a shit. But if I have to pay a dollar, or literally raise my hand, or put in my email or my phone number, then there are stakes.

We have created this system in which all the rough edges are smoothed away. And the reality is you can’t strike a match on silk. That’s not how it works. Friction is the point of flame. Friction is the point of interest.

It’s the point of commitment. It’s the point of most of the good things that come out of life, frankly.

Mike: This is why I cringe every time an agentic commerce demo is booking a vacation! That’s the consummate experience where the four or five hours of “friction” you spend doing research is the enjoyable part. It’s the silliest possible thing to demo.

Jess: One gajillion percent. And the research is all on your side. So much of the joy is the anticipation of the trip and the planning, not the actual trip itself.

Mike: Let’s back up and define agentic commerce because there are so many different definitions living under that umbrella term. What’s yours, as agentic commerce actually exists in the market today?

Jess: My definition of agentic commerce is a transaction that’s completed in a closed loop without a human.

Mike: Ah, so that doesn’t include the entire “zero-click” zeitgeist, where AI is effectively a semantic search layer. Do you think that trend is as important for marketers as the rise of true agentic commerce?

Jess: Honestly, in some ways it isn’t really substantially different than search. It’s worse, but it’s not substantially different.

When we go full agentic commerce and evaluation is compressed into a single API call and all of that is removed from human eyes, and the funnel has completely flipped — that is a different ball game for people and brands.

Mike: You’ve argued that all of performance marketing is a ponzi scheme anyway! Can you elaborate on that a bit further?

Jess: I’m shocked that the Ponzi scheme has lasted for 25 years. Last-touch attribution with promises of multi-touch attribution that never came, never transpired. And we all sort of have been in this collective hallucination that performance marketing dashboards give us a degree of certainty that the brand dashboards they denigrated when they came out didn’t. If Facebook does some holdouts and says you that you got 5% lift, but their dashboard would have reported 64% lift — that’s not even useful. That is actually the opposite of useful.

So that’s what makes it a Ponzi scheme to me. It also takes brand assets and depletes them. All of the work you’ve done to build your brand and your recognition — it harvests it, but it doesn’t place a lot of additional value into the system.

Mike: Has that always been the case, or is it more true now than it was five or ten years ago, across Meta and Google and Amazon?

Jess: Back in 2016, when I was a brand leader at a startup, I started asking our vendor about who we were targeting and how we’d think about holdouts and testing, stuff that was part of my natural inclination having come from performance marketing back when it was direct mail.

And they looked at me like my questions were quaint. Like I was the old dinosaur in the room asking stupid questions. And they were like, Jess, the algorithm is going to tell us that. I was like, huh, fascinating. But the algorithm never really delivered.

The whole point of the Ponzi scheme thing is that it is completely hidden. It is a total black box and people keep handing us receipts that don’t actually add up to anything. And AI and agentic commerce is just going to make that worse, not better. We are going to make even more decisions based on data that we can’t audit, we can’t see.

Mike: But my cold capitalist response is..who cares so long as the arbitrage still works? When will the so-called Ponzi actually break?

Jess: So I don’t know when it breaks, because Wall Street is addicted to it. It’s addicted to the speed. But we’ve got a literal graveyard of brands that probably, if they had spent a little bit more time and money investing in things that compounded for them and actual relationships, might still be around. But who knows, because the projects never got funded or they were deemed failures before the end of the first year.

We never gave brand the grace we gave product.

Mike: So one other place that we’ve disagreed a bit in our banter is around how ChatGPT has laid the foundations for an agentic commerce experience.

I’ve been impressed at how “media literate” they are and how well they’ve been able to parse legitimate information about brands from pay to play slop masquerading as authentic product reviews. I’m pretty excited about this!

Jess: Why should I trust OpenAI to make those judgments for me? As a marketer, as a person — I don’t know whether I trust them to create an algorithm that accurately picks what’s relevant for a brand or for me as a consumer. I don’t know whether those are the right engineers.

It’s not just the data, it’s how the data is constructed, and how we prompt it and how we train it. That’s bias in the data and it’s bias in the engineering. I don’t know whether I want platforms any longer to be the arbiter of anything for me.

And it’s very unclear to me, having literally sat in rooms with Nobel Prize winners, whether we will have the sophistication as marketers to figure out what is going on in those black boxes when the people who invented them can’t figure it out.

Mike: It’s a pretty wild thought– the idiosyncrasies of a couple of engineers end up shaping what hundreds of million people see.

Jess: Yes. And they all look the same, Mike. And that is not a world I want to live in. The more we rely on these platforms, the more I live in a world that is dictated by singular ideology.

And how would I as a consumer even parse ChatGPT versus Claude? The platforms aren’t even honest about how they’re different. They’re truly in brand advertising and not product. I see them telling the brand story, but not at all the product story, and it’s the product story that’s actually interesting to me.

Because it’s not what I use it for. It’s how the answer shapes what I do next. And that’s not what anybody’s talking about in advertising or anywhere else.

Mike: So if the Nobel Laureates don’t even know how these platforms work, my mind comes back to focusing on what you can control. In a world where LLMs eat the world but are going to largely be a black box, what should marketers spend the most calories on?

Jess: What the last 20 years has done is create dependency on platforms, a narrowing in what I can do and what I know of my customers, and a narrowing of the skill sets of the people I’ve hired. We are going into an era where narrative is going to matter more, where above-the-line, IRL things that are actually off platforms are going to matter more.

So many people are getting pressure from their boards to focus on agentic commerce that they’re going to miss the boat on the reality that when that box hits the doorstep, that packaging has to do so much more work. That experience has to do so much more work. That install has to do so much more work. Marketing now has a completely different job.

Mike: And that’s all harder the more you get disintermediated from your end customer….

Jess: But mostly? Own your data. Own your customers. Figure out how to get your fucking customer data. Period. End of. That is the game now.

Sometimes, in presentations, I have this Toast receipt next to a DoorDash receipt. What Toast gives the restaurant is my name — Jess Graham, full name. It gets my full address, my full phone number, and knows I live in Bed-Stuy, which for a pizza place six blocks away is meaningful. Saraghina, by the way. Great pizza. And then I place the same order with DoorDash, the restaurant gets back gobbledygook.

Absolutely no information about me. It was Jessica G. I don’t even go by Jessica. That’s how bad it is. So what are restaurants doing promoting DoorDash? You have another alternative called Toast. You should be doing anything and everything you can to direct your consumers there. It’s the same idea for larger brands.

Mike: Anything in current work you are doing with clients that makes this idea of where you can capture more customer data to own your destiny concrete?

Jess: One of the things that’s not getting much press is the GS1 2027 sunrise — brands putting QR codes and 2D barcodes on packages instead of EPC.

It’s not a mandate per se, but it’s an opportunity for CPG brands and others that haven’t had access to good data. Their data is complete junk. You’re a $100 billion company that doesn’t actually know your customers.

My suspicion is that most of them will use the barcode as a packaging exercise and won’t put together what it could do for the business. They’ll handle it in a silo. We’ve done smart labels and other stuff, and those things never launch with an actual campaign behind it. An initiative.

Everybody’s so busy selling their products. They’re not selling the stuff that could actually save them as an industry. If a new barcode that a consumer can get data off of, and that I can get consumer data off of, is going on every package, you’re going to need to talk about that really loudly and for a long time for it to be worthwhile. Not just how do I use this to sell product, but how do I use it to get data, how do I use it to create stickiness.

Mike: Take the last word. What’s the main message you’d give to a CMO to not be a marketing passenger princess simply wafting on the AI breeze?

Jess: Take some power back. Empower yourself.

I was pitching someone senior at a CPG company recently and I said, well, I don’t think you need to keep the status quo. I think you need to think about your mix in a completely different way and double down on ownership — and double down on the reality that anything you build is going to compound, and anything you rent is going to fuck you in the end.

Jess Graham is the Founder & CEO of Banquet, an advisory practice for the age of agentic commerce. For more of her thinking, check out “Route Around the Machines”, a list of 50 demonstrated tactics leaders can use to avoid the platform tax and reduce discovery debt. To get in touch with her, please email jess@feastwithbanquet.com

To suggest additional builders that I should feature in this series, please get in touch at mike@brxnd.ai

As always, thanks for reading.

— Mike

Purchase Early Bird Ticket

RIP GEO Redditrage, 2024-2026.

Last week, several research reports emerged showing that Reddit, listicles and comparison articles– long the holy trinity of the GEO-industrial complex – were each significantly devalued by OpenAI with the launch of GPT-5.6.

In a sea of reductionist clickbait, Lily Ray has the most comprehensive take on the details that is well worth your time. I agree with the crux of her analysis. OpenAI isn’t trying to curtail the influence of Reddit writ large. Instead, they are trying to disincentivize the practice of brands using grey-hat spam tactics to poison the well before Reddit passes the point of no return. This is the correct approach to ensuring that models aren’t gamed by people who have stacked fake karma or have the best under the table deals with mods.

It will be interesting to see if Reddit comes out the other side of this in a stronger position. There’s a dangerously consensus opinion forming among those close to AI search that we’ve reached the peak of Reddit’s influence over AI, especially as the Google deal looks unlikely to renew. On LinkedIn, I put out a call for some more contrarian bull takes on Reddit which converged on an interesting argument. If they can stop the spammers, Reddit is a constantly refreshing source of up to date content that is “pre-mediated” by both an algo, users and moderators to provide multiple signal points to LLMs on what content is valuable.

Less covered but equally impactful AI Search news is OpenAI also getting fed up with the listicle and programmatic comparison page scheme that many challenger brands have exploited to great effect. Here’s some research from Peec analyzing the prevalence of citation share in both formats the week before and after the GPT 5.6 rollout.

OpenAI is doing right here by its users in making bold moves to limit the efficacy of programmatic slop schemes. But it comes with a major externality. As arbitrage tactics generally mastered first by challenger brands are nerfed, large enterprises will command more of the lion’s share of shaping organic LLM results. The upshot here is for nascent brands, mastering ChatGPT Ads will be far more important than any current GEO tactics. Expect more GEO companies to converge on becoming operating systems for running AI ads.

Broadly speaking, the winds of sustainably influencing models are blowing towards owned content from authoritative brands, creators (especially in AI Overviews where the dynamic of Google leaning on Youtube is only getting stronger) and authoritative, third-party takes from publishers who have deals in place with the labs. In other words, good brand marketing tactics will compound in AI search.

Of these three archetypes, I’d expect to see creators increasingly start to shape LLM responses far more than they do today– much more on this in a future post. The AirOps data below is slightly flawed (given that it largely looks at pre GPT-5.6, it underrepresents how owned brand content is trending currently) but directionally correct on creators commanding far more citation share.

Overall, one pretty safe assumption to make is that by the time the latest AI search hack silver bullet hits your X or LinkedIn feed, any alpha from deploying that strategy is gone. Great marketers generally aren’t in the business of giving away their best secrets on main.

Instinct Eats the World

Last Friday, I got drinks with the most AI-pilled developer I know, the kinda guy where I’ve grown accustomed to sharing his attention with multiple Claude Code sessions that he is concurrently babysitting from his phone. This week, we had a different third wheel, a white hot new AI product called Instinct.

For those mercifully not on VC twitter, Instinct is a personal assistant agent that straddles both consumer and work use cases. The selling point is that it essentially unlocks OpenClaw like functionality for normal people natively through iMessage or WhatsApp, no Best Buy runs on Mac Minis required. In the onboarding flow, Instinct says to “effectively treat me like a human with a computer” and touts the following capabilities:

Instinct is launching into a crowded space that includes other tech twitter darlings like Town and Lindy. There are also more niche consumer agents emerging such as Anna which targets busy parents, whose product leader slid into my LinkedIn DMs as I was doing the Instinct sign up flow. And of course, OpenAI’s marketing team would like you to know that ChatGPT Work can do all of this too!

Or maybeeee not quite. Instinct’s capabilities are wild. Over a Guinness and the worst old fashioned I’ve ever had, I watched my friend’s agent flawlessly purchase several school supplies on Amazon with very nebulous prompts, navigate the rigamarole of a complex insurance portal and surface available Dune IMAX tickets that don’t require spraining your neck to watch the film.

But there’s another side to this story– the early version of Instinct plays…..…just a little bit fast and loose on the security side. Here’s Claire Vo who is an absolute must follow for early reads on any AI product.

The craziest part of the Instinct story so far is that some of the most powerful people in Silicon Valley have handed over unfettered access to their email and bank information to a tool that is in the infancy of shipping basic security features and has a website that looks like this! It’s still unclear to the general public who the hell exactly has built this thing!

If any of this is nefarious, building an agent like this as a psyop to get the most powerful people in the world to hand over sensitive information would be almost too easy a script for a modern day Silicon Valley. The Occam’s razor here is likely a lot less interesting:

1) The investor class that is hyping this up knows the founders well and it is a well credentialed team of AI veterans, not a renegade outfit

2) Investors are so eager to not get disintermediated by a world in which ChatGPT just becomes the default personal and enterprise harness that they’ll pump a promising competitor. It’s not sheer public interest that has traditionally press shy leaders in VC suddenly extolling the virtues of open-source.

I ultimately went through with signing up for an account with Instinct but didn’t connect my email, passwords or any of my work accounts. It managed to snag me an appointment with a primary care provider I’d been calling for months by pinging me the moment Optum released new appointments a few weeks out. But then the weekend beckoned and I didn’t do much else with it.

The next day, I got this text. Presumably, my new friend wasn’t thrilled with my lack of engagement.

Here’s a bit more of our back and forth:

Is this a moment of magic or a creepy dash of black mirror? Yes.

Overall, Instinct has forced me to examine some pretty profound shifts in my own Overton window. One year ago, I would have laughed at the notion of giving any tool like this access to my sensitive data. Now, I came within inches of smashing the YOLO button and likely would have handed Instinct the keys to the kingdom if not for this tweet from Alex Cohen. And here I am, still using the tool in a limited capacity!

A big part of this is a simple utilitarian calculation. As AI can accomplish so much more, I need to rethink my priors around the value I get in abandoning some of my concerns around privacy. But another part of it is a sort of tacit admission that the panopticon has mostly won.

As excited as I am to have an agent do an end-run around institutional bureaucracy, I can’t help but think that this is another example of one of AI’s eternal memes.

Tools like Instinct could well usher in a more egalitarian future where agents replace cumbersome software portals as the access point to most web functionality. Or they could just as easily create a world where a small, privileged group of AI-native elites circumvent the system while the overall quality of ubiquitous internet products goes down. Both will probably happen to some degree in tandem. As more customer service sessions get mediated through tools like Decagon and Sierra, we’ll probably see more companies launch premium tiers where one of the core benefits is that you can always go straight to a person. For better or worse, Cullen represents the sentiment of the populace on this one.

I also agree with Reggie James here. The irony is that as agents eat more of the online ecosystem, more things will run on “knowing a guy” and that the swagless will perish. This is true in both a consumer and professional context and I believe, increasingly in marketing as much as any other function.

Anyway, I have five invites to Instinct to give away. If you’d like a glimpse into the weird frontier of personal agents, drop me a line at mike@brxnd.ai

Consumers Don’t Pay for Productivity Tools

If you listen to one long-form interview this year, make it this Ben Thompson session with Patrick O’Shaugnessy. It’s incomprehensible to me that one analyst could have so much range across topics but Ben galavants effortlessly between AI, geopolitics and game theory analysis in a way that only he can. Most relevant to marketing leaders, he devotes about ten minutes at the 20:30 mark to waxing virtuous about ads, notably reminding listeners of the two core lessons that Silicon Valley learns every decade or so.

1) Consumers do not want to pay for software
2) Consumers do not care about being productive.

Going back to the Instinct discussion above, there are really only two commercial paths forward for most of the “personal agent” tools and most entities building in consumer AI writ large:

1) Pivot to an enterprise / B2B focus and dive headfirst into the gauntlet of competing with ChatGPT Work, Claude and Cursor to be a professional OS / harness.

2) Monetize a free or heavily subsidized paid consumer product with ads.

(One could argue that there’s a third path forward on an affiliate model and in the 2010s consumer tech boom, affiliate models did scale some consumer businesses well into the billions of dollars in valuation. But the cost of new user acquisition was much cheaper back then. There’s a reason Sam Altman was once hot for affiliate economics and quickly pivoted to ads)

I suspect that most large consumer agent upstarts will try both in tandem– the power laws of VC economics will force them to try everything under the sun to hit venture scale. And when they do, there’s going to be a lot of incremental new advertising inventory created.

For a long time, I was skeptical that many of the companies focused on placing “ads in LLMs” outside of ChatGPT would find meaningful scale on the supply side. I’m now much more bullish on this category, which includes companies like ZeroClick (coming soon to a Friday BRXND interview), Koah and Thrad.

This creates a fascinating laundry list of second-order questions! If personal agents and consumer AI tools are going to be monetized via ads, will these be primarily ads directed at humans using the product (i.e. an overtly sponsored response) or ads directed right at influencing the agent at the point of inference? Will people continue to trust their agents if they know that their responses are occasionally being sold to the highest bidder?

These are effectively the same big picture conundrums facing ChatGPT’s advertising ambitions and will largely dictate the pace at which main street gets onboarded into the agentic wave. If AI is going to be anything more than a better search engine to a billion people, it’s going to have to be ad supported.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

Purchase Ticket Here

The AI Search Maturity Matrix

I spoke with a company recently that spends $500k / year on prompt tracking. They track thousands of prompts and there’s an executive-wide email that goes out every week with a “citation rate” line chart that wiggles.

Meanwhile, skyrocketing DataforSEO spend on Ramp suggests that more nimble companies are moving away from enterprise GEO tools and building their own custom solutions in house.

But so what? Prompt tracking isn’t a strategy.

How your brand is perceived by AI models feels (correctly!) existential to many CEOs and CMOs. A core job of the CMO in this moment is presenting a coherent strategy for how to evolve a brand in a world where LLMs increasingly mediate brand preference.

The challenge of building an actual AI search strategy is that you need a way to allocate resources beyond the time horizon of the models, a strategy can’t fluctuate every time a new model drops. The real task, then, is to build AI Search as an organizational capability rather than a dashboard. That’s where the concept of a maturity matrix comes in.

The Five Levels of AI Search Maturity

To say this another way, the core question marketing leaders should be asking is “how can we evolve our AI search maturity?”

AI Search maturity is the organizational ability to turn noisy model outputs into foundational insights about the information ecosystem, then turn that into durable market influence.

To that end, we’ve built an organizational maturity model to help you understand how sophisticated you are - and how you might get better. The model has three dimensions: measurement capability, executive knowledge and strategic posture. Companies are rarely advancing equally across all dimensions.

Hopefully this helps you have a conversation with your leadership team and think about not just problem solving but capacity building.

Level 1: Instrumented - “how visible are we in AI search?”

This is where everyone was 12-18 months ago. You’ve loaded up a set of prompts into a prompt tracking engine and can answer basic questions like “are we more or less visible than our competitors”?

Basic prompt tracking has a number of key problems and challenges though:

• Citations ≠ Mentions ≠ Recommendations - AI search is qualitatively different than traditional search and knowing your ranking position is far less important than knowing HOW you’re recommended. A brand can appear as a citation, but not appear at all in the recommendation or worse, become an ANTI-recommendation.

• You need to aggressively segment your prompts - not all prompts are created equal. Branded vs unbranded, aided vs unaided prompts, different stages of the buyer journey etc. You need to be careful about how you setup your prompt types.

I think more or less every executive team knows AI search is important at this point but at this level you’re still in a kind of vague “change is coming” mindset and have trouble really pinpointing how the new world is different from the old world.

Level 2: Contextual - “which customer moments and sources actually matter?”

Two of the most important characteristics of AI search in my mind is that you can use AI search for more of the customer journey than you could regular search, and that it’s not just ranking answers but actually giving you qualitative advice.

So your setup has to grapple with this reality with:

1. Prompt segmentation that maps to your customer journey. The good news is that most brands already have a reasonable notion of their customer journey from brand marketing - so the first step is to map your prompt tracking to that journey and to use the same kind of language the brand team uses (aided vs unaided awareness, likelihood to recommend etc)

2. Qualitative analysis of the prompt responses. Beyond just knowing how often you show up you need to be probing the responses for qualitative insights - are we being recommended strongly? Which gaps exist where we’re being cited but not recommended?

At this stage your leadership should be able to articulate gaps - a difference in how the brand shows up in AI search vs regular search. You should be able to point to areas of weakness and strength.

Level 3: Operational - “Can our organization act on what we learn?”

Once you know where the gaps are (level 2) the next level is acting on them. Most AI search problems are consensus problems across the internet. Do we have a reputation and brand awareness that’s strong and repeated everywhere?

The strategic work becomes coordination. Brand, content, product data, PR, community and technical teams organize around the information gaps the company has found. Some gaps require a new page while others might require better product data, a clearer position, stronger third-party evidence or fixing the underlying customer experience.

Citation rate and mention rate varies tremendously in AI search but measuring consensus is a little more long-term. To do that your measurement stack has to be able to probe the qualitative prompt responses not just for sentiment but for coherence: are our brand attributes correct everywhere? Is it mentioning the brand values that we want to repeat everywhere?

It also starts mapping the information ecosystem producing the answers. Is the model drawing from the company website, a retailer feed, Reddit threads, review sites, old press coverage, a product database or a competitor comparison page? Where does the ecosystem agree with the company’s intended positioning, and where does it tell a different story?

In my experience very few companies are actually at level 3. Most are at level 2, aspiring to become level 3.

Level 4: Exploratory - “Are we paying attention to how AI search is changing?”

Despite all of the hype about AI search, I still don’t think people appreciate how much better the models have gotten. AI has crossed the chasm of building a truly superior search experiences for everyday consumer use cases than ten blue links. Recently on vacation I had ChatGPT using skills, operating on scheduled actions and providing a level of recommendation and usefulness that regular search can’t touch. I used Google Maps a bunch but I can’t recall using Google once.

Getting cited and mentioned is one part of AI search today but I think you should be paying attention to the frontier too. Click based analytics is effectively dead vaporware. Brands can be discovered and recommended without a single site visit. Agents can autonomously research and purchase products.

No one is ready for this world but smart brands should be investing in exploration. Some areas that I might invest in here today:

• WebMCP:- While agentic commerce protocols still feel early, clunky and not widely used, the idea of giving every website a structure, agent-native access point will really upend a lot of things. It feels like we’ll see some major innovation here in the concept of a webMCP before the end of the year. (one to watch: Cloudflare can now give every site a webmcp endpoint)

• Model memory vs agentic web search - beyond mention/citation rate, it’s important to understand how the models understand your brand and your products in their training data vs what they learn from agentic web search. Open source models can provide raw thinking traces and can shine a light on various brand attributes associated with your brand and allow you to probe model memory more directly

I only know a small handful of people operating at this level.

Level 5: Experimental - “Can we prove that our interventions work?”

Level 5 is the crown jewel. Can you actually test and iterate on AI search as a field? This is harder than it sounds given the volatility in the space.

There are two promising pathways I’ve seen here:

• First, using a platform like SearchPilot (disclaimer: my brother runs it) you can actually construct A/B tests for site changes to understand what moves the needle. True for regular SEO and increasingly for AI search too.

• Secondly - you can build simulators that enable you to actually feed different context into the models and see how they react. This allows you to estimate the impact of changing owned or influenced properties and seeing how it changes the likelihood to recommend.

We’ve been playing with this approach to build simulation environments and see how changing descriptive language changes the likelihood to recommend and whether brand / product features influence the models.

Here’s a glimpse at the output - we took a real hotel and ran hundreds of simulated AI recommendation queries against four real competitors, changing only the wording of the listing each time. Reordering the copy to lead with what business travelers care about took the model’s recommendation rate from 16% to 78%, while removing the brand names barely moved it: the words, not the brand, were driving the choice.

More on this in later newsletters but get in touch if you want to explore this for your brand!

Where is your brand on the maturity matrix? Where is this framework wrong? How would you extend it? We’d love to hear from you.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

But first, BRXND NYC is now 80 days away on 11/5, and tickets are currently $749. I look forward to seeing you at the Times Center for a great show.

GET YOUR EARLY BIRD TICKETS

In 2023, I visited a high school in LA to talk about AI. It was the early days of ChatGPT, and mostly I answered questions about how it worked. After speaking with the students, I sat down with an English teacher one-on-one who was worried about what it meant for her and her teaching. My answer, which I’ve come back to a lot over the years, is that in theory it shouldn’t mean anything: her job was never to make great writers out of 15- and 16-year-olds, it was to convince them that learning to write and getting better at it was a worthwhile pursuit. Fully respecting that I’m far out of my depth as a high school teacher, I’d argue that, in theory, AI shouldn’t change that job.

Fast forward a few years, and I’m sitting with my fifth grader, helping her do her homework. It’s the first time she’s being asked to write paragraphs, sometimes more than one. These “write longs” are graded assignments, and she wants to do well. I offer to work with her on it, trying my hardest to find that proper parental line between guidance, advice, and doing the work. Her natural approach is to read the example her teacher provided and try to adhere to it as closely as possible. This makes my advice easy: don’t do that. We talk through what the assignment means, and what the example represents, and then I encourage her to just tell me what she’d like to say. After that, I encourage her to write it down.

The thing about writing, good writing at least, is that it should sound like it could only come from the person that wrote it. What good writers have is a confidence in their own voice. Sure, they should know the rules—at least the basics. But honestly, the rules matter less than having something to say and finding a way to say it that is true to yourself. Sometimes you know that ahead of time, but sometimes you only find that out by writing and re-writing. The challenge with AI in writing, and just about everything else, is that it’s training as a “helpful assistant” and aims it towards the end product. It always wants to deliver the document or write a ticket, when often what you need is to think through the problem. I have always found the best role of writing is externalized thinking. It’s also why I hate conclusions.

One of the most confounding things about the recent releases of even more powerful AI models is that they seem, in my view, to be getting worse at writing. I suspect part of it is that the explosion of AI for coding has driven more and more focus on verifiability as a key measure of quality, something writing clearly lacks. I also think, to the point above, that reinforcing the “helpful assistant” voice has probably driven a certain flattening of language. But there’s also another side of the problem: there are certain things like em dashes or “it’s not this, it’s that” that used to be perfectly fine and are now a danger zone because LLMs have converged on the pattern. When you saw an em dash above, you might have had a moment where you thought I had used AI to write this (I haven’t).

The problem, I think, is that good writing is fundamentally anti-consensus. For writing to be above average, it shouldn’t sound like everything else. My daughter might get a fine grade aping the teacher’s example essay, but the road to good, or even great, is doing it your own way.

On top of that, writing is a creative act and, as a creative act, the process matters as much, if not more, than the outcome. There’s a saying in law I was introduced to years ago: “Does it write?” Writing is figuring out whether there’s a there there. It’s an opportunity to test the validity of your thesis on yourself: can you string words, sentences, and paragraphs into an argument that still captures your intention and attention 700 words later?

With all this said, I’m obviously an AI maximalist. I used AI to collect thoughts, pull together references, and think through this essay. I believe we will and should use AI in certain kinds of writing. But I also believe that AI writing is fundamentally not good at the moment, and you should feel embarrassed to ship slop to readers. Like a lot of hard problems, there’s not a simple way to draw a line. But I think some good rules of thumb have emerged:

1. Writing is thinking, and if you’re interested in pushing forward your own perspective on an idea or problem, you should be the one hitting the keys.

2. Every sentence matters, and you should be prepared to stand behind each of them, regardless of whether you or AI wrote it. Respecting your reader means spending more time writing something than it would take them to read. (If your reader is another AI, then aligning to token time is obviously appropriate.)

3. Never let AI remove you. We are going to use lots of AI. I ran this essay through a few different systems (spelling, grammar, ideas). We should keep doing that, but it’s also okay to say no to suggestions. Grammarly wants me to add a bunch of semicolons and fix some technically incorrect sentence construction, but semicolons are not my style, and using a few extra “reallys” sounds more right in my head, so they’re staying put. If you don’t feel certain, it’s okay to have it read uncertain even though the AI and the writing manuals tell you not to. That’s you.

4. Slop kills good ideas. Sadly, even if you’ve got something great to say, if you open with six slop tells you’ve lost me.

5. Good writing is vigorous. Strunk and White said it best, “Vigorous writing is concise. A sentence should contain no unnecessary words, a paragraph no unnecessary sentences, for the same reason that a drawing should have no unnecessary lines and a machine no unnecessary parts. This requires not that the writer make all his sentences short, or that he avoid all detail and treat his subjects only in outline, but that every word tell.” Sadly, most LLMs treat writing like college freshmen: length is the goal.

6. Writing is anti-consensus: your voice is what makes your writing good, and no matter how much prompting you give a model, it’s not currently capable of getting there without real help.

7. Becoming a better writer is a worthwhile lifetime pursuit. None of us has reached the top of the mountain, there are always places to improve and new words to learn. We can always communicate our ideas more effectively.

This post was in part inspired by Clay’s writing policy and Vicki Boykis’ Write for people post.

Since you’ve made it to the bottom, one more reminder that BRXND NYC is coming up on 11/5, and tickets are currently $749. Get yours early, we will sell out.

GET YOUR EARLY BIRD TICKETS

As we build towards this year’s BRXND NYC conference, I’m continuing our series of interviews with founders, technologists and marketing leaders innovating in the problem space around how AI is fundamentally reshaping the way consumers build preference and purchase brands.

If there’s someone you think I should get to know, please drop me a line at mike@brxnd.ai. If you’d like learn more about our November 5 show at the Times Center, check out details here or click the button below to grab an early bird ticket.

Buy Tickets Here

On Content vs. Knowledge for LLMs

Much of the modern web is a corrupted morass of arbitrage hacks from a bygone era that have rendered media utterly impossible to consume. You feel this most viscerally when you look for a simple chili recipe and have to read 5,000 words about how Grandma Ethel survived dysentery on the Oregon trail before you even get to put the onions in the pot.

That mess is now the raw material feeding every LLM on earth, which is why the machines spend so much of their energy sifting five thousand words of Great Plains hardship to find "one large onion, diced." For Alex Springer, this is the largest opportunity for CMOs. AI systems don't want to consume content, they simply want the knowledge buried inside it. And the vast majority of knowledge about products is owned by brands directly.

Alex has spent a decade building marketing, measurement and reporting teams and technologies across the US, UK, and Europe. His primary focus is transparency, accuracy, and accountability in AI systems. He serves as tech lead for the SPUR Coalition of publishers (founding members include the BBC, Guardian, Telegraph, Sky, and the FT), is Director of OpenAttribution.org, the open standards body for content attribution in AI systems, and co-leads the AI Taskforce for the UK based APMA.

He is quite simply the most erudite man I know on how AI agents intersect with content and my go to WhatsApp sparring partner for dive bar conversations held across an ocean on all things about how LLMs view the world.

A look into our general Tuesday morning DMs and lightly edited version of our conversation follows below:

Mike: Alright Alex, Other than pedantic nerds like us, who’s actually digging into what makes an agent pick one product over another? And what are they finding?

Alex: So the team working on the UCP (Universal Commerce Protocol) at Google set up this lab where they’re loading in a bunch of different product catalogs and running experiments to figure out what makes agents buy what. We are still at this rudimentary level of saying, okay, here’s ten laptop chargers, all features being equal - which does the agent buy? The cheapest one, obviously (spoiler: this doesn’t bode well for brands charging a premium based solely on brand value).

But there were interesting edge cases where say nine of the chargers had amazing features; however, the “lesser” charger had better safety ratings from a registered safety organization - and the agent always bought that. I’d say this makes understanding the bias of a model - all of the big frontier models are tuned to prioritise safety - a critical thing to understand.

I was at a talk from one of their engineers and he was talking about this “quirk” - I pulled him aside afterward and said “I know you said that as a throwaway statement, but marketers are so desperate for this information that whole companies are raising millions on much shakier intel.”

Whenever Google’s engineering teams meet up with their biz dev folks (won’t be long) and get these insights into their ads and search products it’ll knock out half the GEO offerings out there. Until then I suppose the GEO LinkedIn-industrial complex can make a few hundred grand on consulting engagements based on these little hacks?

Mike: But do these safety ratings even have to be real? Or is this another case where we can throw “eco-friendly” or “organic” on everything and fool the models?

Alex: Well, that’s the right question. Because if I throw out official sounding acronyms, most of the LLMs will just be like, “oh yeah, sure, that must be a thing.” It’s a probabilistic token predictor going, yea “ABCA-certified” does totally sound like a real thing, why not? Don’t believe me? Pick an even slightly obscure acronym and ask AI Overview about it.

Here’s Google’s response to the UK CMA when it was suggested that its existing mechanisms regarding factuality were not good enough:

“Google receives … feedback via the thumbs up / thumbs down mechanism every day. These simple visual icons allow both publishers and users to quickly flag if something is incorrect, with a specific option to select ‘not factually accurate’.” Link

“Publishers” in this context is anyone whose content is being scraped and regurgitated. I’d bet 99% of brand marketers and 100% of consumers didn’t know that it was up to us to help double check every answer on Google’s behalf.

So now we have these answer engines mediating our access to information without being held accountable for incorrect statements by any of the consumer protections we in many ways took for granted. We need to think about how you signal authority in a verifiable way to anything that gets injected into the agent that’s doing the stuff. And that’s the tricky part.

Mike: Ok, let’s zoom out a bit. Influencing LLMs is obviously a Herculean abstract challenge…and a whole cottage industry that is less than two years old.

Who has the best framework for where a CMO or head of AI transformation can truly dive in and start putting tangible points on the board right away?

Alex: I recently had a great conversation with the chief AI officer at an international, multi-brand retailer - someone who is also deeply involved in broader industry standards work. He’s basically going piece by piece through their brands saying: marketers, get your house in order, and the first thing you do is align product feeds and owned and operated site content.

Forget the third-party content for now. We can’t control that, it’s messy, the LLMs are changing the algorithm every day, and we’re not going to get into the SEO battle again. They don’t care about that, they care about our product data being well organized.

It was the same conversation I had a bunch at Cannes this year. I’m a brand: the first thing I do is make sure I’m clear, consistent, available and accessible with my data. Then I can go try to mess with whatever third party wrote a good or bad thing about me.

Particularly with the sea of content laundering and shitty arbitrage URLs trying to pump authority by not blocking while premium publishers block and negotiate better deals with the models. It’s such a mess in there that the CMO looks at it, goes a little cross-eyed, and goes back to what they can control.

The first question I ask a CMO? What is it you’re known for? How do you describe your brand/product/service? Then you go ask ChatGPT the same question and do the hard work of reconciling the inevitable disparity by tracking the inconsistency across the web. Because the LLMs don’t have opinions - they learned to talk about your brand from somewhere.

Mike: “Fix your feed” is a pretty anticlimactic headline man…….

Alex: People look at me with a lot of disappointment when I say that. They’re like, “but that’s not clever! I thought listicles were the answer!”

LLMs have exposed that all of our data is really messy, globally. And LLMs are going to force us to finally fix that, instead of pretending we’re going to fix it. The only way to do that is to finally do what we said we would do for years and get our data in order.

You want a slightly better headline? If your brand info and image isn’t accessible, accurate, consistent, and observed - forget about how the answer engines talk about you. That is just a symptom - the root cause is closer to home.

Mike: So how can brands make sure they are marching in a direction that is long-term aligned with how the models want to consume information vs. playing short-term arbitrage games?

Alex: I think doing stuff like using LLMs to blast LLM-written content about your brands across Reddit will bite people. It already is — they’re downranking LLM-written content left and right. The temptation to do it at scale is probably the biggest gotcha. And Reddit’s deal is expiring soon. Looks likely enough it won’t renew so Reddit’s not going to be ranking for much longer either.

There’s also a poisoning-the-well issue. All the model builders I’ve talked to are pissed off when they realize — sweet summer children that they are — that half the internet is biased, because we’ve been biasing it forever via advertising. If models are going to tell the “truth” they need consistent and ideally true information, and the internet is full of paid-for, semi-false information. The more that’s true, the more they run away from various sources.

Too many brands are putting a lot of short-term pressure on third party publishers, basically paying a lot of money to write a whole lot of stuff that caters to the brand. Increasingly that’ll be cross-checked, and increasingly the models will back off of it which hurts both brand and publisher in the long run.

Mike: Oh man, I played this game hard when I was a growth leader two years ago in Google’s glue on Pizza era. Pumped out both owned programmatic SEO and worked with basically glorified content mills to get pretty small DTC sites showing up in AI overviews for “best x for y” searches. But, it always seemed like an ephemeral thing.

Alex: Yup, the original version was “just suck up everything”. Now there’s more of an ethos of “what is the least amount of stuff we need to ingest that teaches our model?” And then, how do we get really, really specific about high-quality inputs?

Mike: There’s a bit of a contradiction here though. LLMs clearly go to third parties because they know they can’t trust what brands write about themselves without some kind of independent validation.

Alex: Yes… but think about what is the most authoritative thing about a product that actually has any kind of accountability. It’s the information provided directly by a brand. You shared something with me from Lily Ray the other day about that right?

Mike: Yup, GPT-5.6 is increasingly going directly to brands and manufacturers to retrieve product, feature, specification, and pricing information

Alex: I trust brands talking about themselves, because I know that there are consumer protections if the brand lies to me about itself. If a publisher review fudges some numbers or goes out of date, there isn’t necessarily the same recourse.

And there are areas where the models intentionally still get things wrong. The system prompt for most major LLMs at the moment includes something along the lines, multiple times, of don’t quote from more than one source more than fifteen words in a row. Because if you do that, we suddenly get hit with copyright infringement lawsuits. They haven’t gotten worse at being able to reproduce content at the model level!

There’s no accountability mechanism for a lie, but there is an accountability mechanism for theft. So they sometimes err on the side of lie.

No brand is ever going to sue Google for using their Merchant Center product feed data. So the model can go, oh cool, yeah, you can quote that directly, that’s great.

Mike: So how does a brand-side marketing leader start to actually own shaping the narrative AI weaves about their products beyond just exposing accurate information?

Alex: A lot of the work I do is trying to empower publishers and brands to own their own content and serve it out in smarter ways, as opposed to just handing their content over the platforms and hoping that their general models serve as decent sales reps.

At core what AI systems need isn’t content - it is knowledge. The current system is super inefficient because for decades we’ve been really good at hiding the knowledge inside of ever longer content (see the damn recipe blogs that no one wants to read - recipes are too short to serve a bunch of ads on though!). And right now the only recourse really is for the AI system to take it all and then do a lot of processing, sifting for relevant knowledge.

Mike: There’s some irony there in that I’d argue that “hiding knowledge inside of ever longer content” is the fatal flaw in LLM writing as well…..

Alex: But if we start thinking instead, I as a CMO own my own content, own my own IP, and I can enrich it how I want and experiment with ways to make it more attractive — I can turn my content back into knowledge!

For a long time agents have had more immediate concerns: “how do I make sure I’m not getting prompt injected, and how do I not blow up my token count?” That’s finally evolving into, “how do I put stuff in my context window that’s going to create better outputs and outcomes.”

That is the real opportunity - and you can see it in nascent efforts from CDNs to make “markdown versions” of the web, in content grounding services serving “chunks” instead of articles… but who better than the original content owners and creators to offer their distilled “knowledge as a service”!?

It’s scary out there and things are moving fast - whatever form this new knowledge economy takes I think it’s a safe bet that the marketing leaders that own their own knowledge will do better than those who give it all away and hope for the best.

Mike: Does this imply that brand-side CMO’s have more leverage with LLMs than they’ve been exerting? AI providers need up to date brand-side content to best serve their users! What should brands ask for in return?

Alex: If you’re a FORTUNE 500 advertiser right now, you say: look, OpenAI, we have billions in advertising I would happily put through ChatGPT. I cannot do that as long as I’m in the dark about how you’re targeting. And I can’t do it if I don’t know whether you’re saying hand lotion causes hairy palms and then……boom here’s my Neutrogena ad!

I think that the leverage has to be tied to ad spend rather than any notion of withholding access to content as publishers are doing. Can’t imagine many CMOs who are going to say, “oh, I’m going to take my Merchant Center feed away from Gemini.”

That said - your product data is valuable to these models. Which does give a brand leverage.

Mike: Can you get a bit more concrete– if you’re a CMO with the power to spend nine figures in ChatGPT Ads, what do you ask for concretely from OpenAI before you move forward with a large ad buy?

Alex: Citation and grounding manifests for conversations in which my brand is listed - you have to be able to know what content (yours and others) was shown to a user, what informed the answer.

Really understanding grounding—which is to say deeply knowing which content from where is influencing the agent at time of inference — is the main rallying cry on my side.

This isn’t just a brand safety issue from the consumer point of view - media companies are getting aggressive and suing when their content is used without license. Not just the grounding services but the agent builders using those services - why wouldn’t they happily sue the advertisers that are paying for the content theft as well?

If we don’t know the inputs for an AI system, we can’t trust the outputs.

Mike: What about an ad placement next to an organic prompt response that says you shouldn’t buy a given product at all?

Alex: The thing that should scare performance marketing types most is that if I say I’m going to go buy some golf clubs, the first thing the LLM often says is go buy used. I have to practically beg ChatGPT to suggest I buy a new car. It wants to give me tips on making my old one last.

That should scare the shit out of the entire industry, because its priorities are not to go sell me new stuff. Its priorities are to save me, as the end user, money. Figuring out how to efficiently deliver ads that drive any kind of engagement and real action — in the face of a machine that is trying to more demonstrably serve the user— that is causing an ad tech crisis, to say the least!

Mike: It will be really interesting to see if this dynamic holds as OpenAI really hits the throttle on their ads business or if the training is tweaked to make the underlying environment more ad-friendly. OpenAI swears that will never happen….but a year ago Sam also swore they’d never run ads!

Alex: This gets back to the biases trained into the models and harnesses. They don’t “think” buying used is inherently better, or that climate change is real, or that democracy is good - they have been tuned that way. Again we need visibility - otherwise it is just a couple of key presses from nameless engineers and we won’t know for ages that our “thinking partners” now think something very different.

Mike: Take the last word here, how is the work you are doing both at OpenAttribution and with the SPUR Coalition helping to provide more transparency in how LLMs form their opinions and notions of brand?

Alex: I firmly believe that we as consumers, marketers, writers… humans deserve to know why a generative AI system believes what it believes and responds how it responds. Everything else proceeds pretty naturally from that premise - and from that transparency we could build a functioning knowledge economy that benefits all sides (including the model builders).

SPUR and OpenAttribution give us the opportunity to convene and find common cause around that concept. The media companies that form the SPUR Coalition have a very immediate need to solve this problem - AI platforms are substituting their content products, denying them user visits, and in most instances refusing to engage in constructive solutions with the broader industry. Industry coalitions can and have been a useful trigger for that engagement. OpenAttribution offers a similar convening across the brand and advertiser space.

We are working on pragmatic solutions - we’ve set up a suggested common language for how AI systems ingest and use content and data as part of their answer generation process and we engage with AI builders, brands, publishers, and regulators along with other standards organisations in order to promote transparency and accountability in content usage. This transparency directly serves brands just as much as anyone.

OpenAttribution also works directly with brands to assess and improve their posture is taken to AI - data availability, bot management, etc.

I’d throw out an invitation to your audience Mike. Brands can either wait it out and hope that the content producers win their battle and just take whatever data Google, OpenAI, etc deign to hand out - or they can recognise that their clout, money, and connections give them a say.

Brand safety, accuracy, accountability and ownership should still matter - even if for this moment, many seem willing to pretend like they don’t.

Alex Springer has spent a decade building marketing, measurement and reporting teams and technologies across the US, UK, and Europe. His primary focus is transparency, accuracy, and accountability in AI systems. To get in touch with him, please email alex@openattribution.org.

To suggest additional builders that I should feature in this series, please get in touch at mike@brxnd.ai

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

Purchase Early Bird Ticket

Does the opinion vibe?

In the last few weeks, I’ve finally gone headlong into vibe coding, embarking on a poor man’s version of the journey Ben Thompson has also been on. Along the way, I’ve built:

1) A leaderboard of GEO companies, which I wrote about in vol. 122

2) A job board for forward deployed engineering roles, the hottest gig in tech.

3) A gift finder for indie products called Almanac. This is made possible by the Shopify Catalog API, an incredible “arm the rebels” offering from Tobi and company that I highly recommend anyone passionate about commerce plays around with.

While these aren’t even low A-ball numbers among my Alephic peers, I’m proud of this little chart below:

As I’ve experimented, I’ve found the cerebral challenge around vibe coding to be frighteningly similar to the agony and ecstasy of writing.

The misery of writing is that it exposes critical flaws that can be papered over when an idea is just an internal monologue in your head or spoken aloud in a board room. Tautologies masquerading as substantial ideas get viscerally exposed the moment they are manually put to pen.

Paradoxically, when done right, vibe coding is much more like the arduous process of writing manually than it is like generating text with an LLMs. Even though AI does the work of writing code, building even a basic application forces you to engage thoughtfully with what it is you actually want to build in a way that even most PRDs never did. When you ship pure slop, you can feel deep in your bones that the root cause is that you have no idea what you even set out to build. And more than ever:

In a marketing context specifically, the most interesting thing about vibe coding is that it shows you where the unique alpha in your data truly lies. Much of the SaaS-industrial complex that CMOs have relied on to date boxes you into a commoditized way of working built around their lowest denominator customer. Wrestling with the nuance of building even a basic vibe wireframe MVP of a software product you’d generally buy helps you deeply understand how tech should serve your brand, not the other way around. If your brand is known for its irreverent wit, it’s kinda crazy that you’d just use the same social media management SaaS as a bank!

So here’s my challenge to any non-technical marketer reading this newsletter. Before you consider purchasing any solution, ask yourself, “does it vibe?” By this I mean spend two hours in Claude Code or Codex describing the core problem you are trying to solve with a piece of software (or for that matter, an agency partnership).

Verbatim enter into the chat the precise problem space that is keeping you up at night. Here’s an example staying close to the GEO space:

“I am a marketing leader at a fast-growing wellness brand that sells products in multiple categories. While our brand is often cited by LLMs, recently, I’ve noticed that ChatGPT is making several factual errors when I run queries that surface my products. In other cases, information is correct but I don’t care for how my brand is perceived. Please help me design a lean solution to proactively surface factual errors, and places where LLM descriptors of our brand differ from the copy in our owned channels. Even better if we can build workflows where agents proactively seek to solve these problems once identified. Before you build anything, please ask me any clarifying questions to help design the best possible solution for my brand.”

The prompt below is a sort of rough amalgamation of a problem space (more on this in an upcoming post on The GEO Maturity Index) that Profound, Bluefish, AirOps and other leading GEO solutions have all recently launched products around. At first blush, it might be hard to nail down which would be the best fit for your specific use case. But my actually starting to vibe out the problem, you can at worst be a much more discerning buyer of software. Ultimately, you might even find there are more custom things you need to push a partner to build or just start to own yourself.

Who knows if vibe coding will ever emerge as a legitimate means of replacing core business infrastructure– for now, no sane marketing leader should attempt to rebuild Hubspot or Klaviyo from the ground up. For more bespoke builds, you certainly want professionals on the keys shipping any production level code. And for now, vibe maintaining software is a hell of a lot less fun than vibe building.

But the world is changing fast and fortune will favor the executives who spend $200 hitting the limits on a Claude Max plan in lieu of empty pontification. Those who embrace vibe coding will find themselves increasingly leading from the front and getting to velocity faster on their ideas by “seeing if their opinion vibe codes” alongside their teams. Like writing, vibe coding is thinking.

Cloudflare and Stripe Enter the Marketing Arena

One of my wackier face value takes is that Cloudflare and Stripe will be two of the five most important companies of this era for marketers, joining Google, Meta and OpenAI. My thesis here is pretty simple– if you own the (increasingly agentic) transaction layer of the internet, you will become the arbiter of how media is spent.

Cloudflare in particular has been on an impressive tear, last week launching a webMCP interface, agent-first browser, and open source OS as part of its agents week. Broadly, speaking I agree with this take from Daniel Miessler about how much they fly under the radar relative to the beachhead position they occupy in AI.

Most pertinent to marketing leaders, Cloudflare also unveiled an “AEO Visibility Dashboard” as part of its agentic readiness kit. Credit to Carl Hendy on X for giving us a first look at the interface.

Cloudflare executive Lara Cohen wasn’t shy in positioning the offering against the status quo in the space. This is a slightly unfair quip, but one that will resonate deeply with many marketers nonetheless.

So what’s Cloudflare’s game here? Why launch the 30th tool in this category? Basic GEO / AEO prompt monitoring in and of itself is a commodity with limited utility but as part of a suite that helps turn websites “agent native”, it uplevels into a larger narrative around “a new business model of the internet.”

In the abstract for this year’s BRXND conference, Noah wrote “there is a not-hard-to-imagine future where the most important job of a marketer is to convince the model it is worthy of its user’s attention and dollars, and, even further, where the LLM just has its own credit card and makes the decision on a consumer’s behalf.”

Cloudflare is effectively aiming to be a full operating system for brands navigating this future state, an extremely lucrative lane and a potential welcome dose of ibuprofen for marketers currently navigating disparate technologies across AI search and agentic commerce readiness.

Deeper in the throes of Twitter, Stripe also seems to be experimenting with features aimed towards marketing leaders. From Nigel Yong, here’s a snapshot of attribution built directly into a Stripe subscription page.

This isn’t as random as it seems! Dating as far back as last fall, one of Stripe’s product leaders put out a call for design partners who spend heavily on search and social.

Now, it looks like they’ve started to build something that should make many technology companies nervous and marketers pretty intrigued!

Much like Cloudflare’s GEO play, if this were just basic attribution of Google and Meta campaigns bundled into payments reports, it isn’t all that interesting. But I can’t imagine Stripe would care to stop there. Their core mission is to help merchants grow faster (or increase the GDP of the internet in Stripe parlance) which in turn directly increases payment volume on Stripe’s platform. As the machines eat media buying, I reckon Stripe wants a piece of that action to increase the share of merchant activity where they add value.

Stripe has an interesting right to win with marketing leaders if they build an agent that can buy ads. They see the actual revenue, recurring payments, refunds, churn, and true customer lifetime value of transactions from merchant payment data. Who knows what will become of click-based conversion signals? The payment layer is eternal.

Both Cloudflare and Stripe have significant access to first-party data—directly at the point of transaction and now inference—that a generation of software companies mostly took an educated guess at. They are, to borrow the famous turn of phrase uttered to the Google founders by Viacom CEO Mel Karmazin, once again “fucking with the magic.” Only this time it’s a lot of legacy middleware SaaS rather than the media industry in the crosshairs.

A few more things on my mind….

OpenAI wunderkind Gabriel Petersson launched a new company called Energy that appears to mostly be ChatGPT for Work / Codex wrapped in a more user-friendly interface. At first glance, it seems primarily like a narrative arbitrage for people who don’t already know what existing harnesses can do? And it might work?

In our last newsletter, Noah’s first “things he thinks he thinks about AI were:

1. Everyone is going to be working in harnesses (Codex, Claude Code, etc.) in the future.

2. I’m not sure if that harness-first future is 6, 12, 18, or 24 months out.

I agree wholeheartedly on both fronts. And I wonder if there just may be a weird window between mostly browser based work now and when everything just happens in Codex or whatever enterprise harness where little upstarts like Energy can sneak in and build incredibly valuable companies.

Per Digiday, Perplexity has blocked Time’s ads served to AI agents, calling them “deceptive.” The framing here is a little rich from Perplexity, considering that the company was the first AI search platform to enthusiastically roll out ads back in 2024. When their efforts stumbled, they became crusaders of the sacrosanct, ad-free AI-search experience.

That said, it does expose a fundamental issue of incentive alignment among different parties in the “selling ads to agents” business model. Foundation models and other AI search platforms will understandably be hesitant about any offering that makes it feel like there are easy ways to influence their search index! At the same time, agentic search will ultimately dwarf human web search in volume and it’s pretty naive to imagine this will all happen without an advertising layer baked in.

Like many other things around AI, this is all so early.

Gartner’s “Magic Quadrant for AI-Augmented Code Modernization Tools” is the corporate performance art piece of the year. Someone tell Dario and Elon they are cooked.

Tim O’Reilly penned a fascinating response to SemiAnalysis’s take that Google is no longer a frontier research lab, evoking George Westinghouse’s bet on AC generators to diffuse electricity through society. Both pieces are well worth a read.

The Onion continues to thrive in its rejection of AI and embrace of physical newsprint, the only company I can think of that has found success in the “artisanal, anti-AI” brand positioning.

At some point, you have to wonder if a declining, ZIRP-valuation SaaS business with few other options is going to take a stab at this counterpositioning playbook. Reject all AI-first and conversational chat experiences in their product and position their tool as the easel for people who love truly “doing the work” and building things themselves with software.

It will be near impossible to pull off– even Canva’s power users are decamping for ChatGPT Images. But for a founder at the brink, it may be more fun going out in a blaze of glory than having their silverware contorted by a group of Italian private equity barons on the way down.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

But first BRXND NYC is 90 days away on 11/5, and early bird tickets are currently $749. Get yours early, we will sell out.

GET YOUR EARLY BIRD TICKETS

Things I Think I Think About AI (2026 Edition)

1. Everyone is going to be working in harnesses (Codex, Claude Code, etc.) in the future.

2. I’m not sure if that harness-first future is 6, 12, 18, or 24 months out.

3. If your primary experience of AI is still in ChatGPT, Claude, or Copilot, you have no idea what these models are really capable of.

4. Don’t count out Google and Microsoft in the harness race quite yet. I still think the fundamental question is whether these model providers can build great interactive surfaces before the interactive surface providers can figure out how to properly integrate models.

5. Anyone who says AI can’t do useful work is lying to themselves or to you.

6. Fable is the best model on the market, and I don’t think it’s close.

7. Fable is the first model that feels smart enough that you shouldn’t waste its time having it write simple code. Its best use is as a planner.

8. There are only three real vectors to judge models: Raw intelligence (frontier foundation models), cost-per-intelligence (value models), and tokens-per-second (there’s some minimum bar for intelligence here, but it’s low), everything else is competing for scraps. The frontier foundation models are where we all pay the most attention (Fable, Opus, Sol, etc.), but I think the most interesting/dynamic space right now is cost-per-intelligence, where you have Gemini Flash-Lite, 5.6 Luna, and open source models like DeepSeek Flash v4. These are workhorse models and play a huge role in almost any large-scale AI project.

9. I still think the hype around local models is far from the reality. The only thing I find myself using them for at all is a tiny bit of local search/embeddings stuff and almost nothing else. The good open source models are too big to run locally, and everything is just too slow to make it worth it (before you ask, I have an M5 Max with 128 GB RAM and a DGX Spark in my home rack).

10. Codex Desktop (now ChatGPT) is awesome and lowered the barrier to entry on harnesses in a really powerful way. I don’t understand why everyone isn’t using it all the time.

11. The tokenmaxxing conversation is stupid: the vast majority of companies should be spending way more on AI, not way less.

12. The best way to learn AI is to buy a $200 Claude or OpenAI plan and try to exhaust your limits.

13. There are a lot of companies that are trying to optimize their AI spend before they have any. That is a dumb mistake that will come back to bite them when, in a year, they still have no one using AI.

14. While token leaderboards are a blunt object, generally tracking adoption and making sure teams are using the technology is a good thing to do. Obviously it can create bad incentives, but the more damaging incentive is individuals doing nothing.

15. The number one challenge with using AI to write code isn’t buggy code, it’s misalignment with vision and architecture. These models are already just as good (if not better) than your average mid-level engineer, but they’re terrible at deciding what’s valuable to work on and staying in line with the technical architecture that’s been laid out without quite a bit of coaxing.

16. There are two primary modes for using AI: divergent and convergent. The latter is for problems where you know what you want, and you should be able to just run it through some kind of system and let the AI take care of it. If it does a bad job, it’s almost always a failure in how you designed the system, not a failure of the AI. The former is how you use these tools to think through new problems and ideas, and it’s a much less well-explored space. Here you don’t want to give over judgment to the model, you want your taste and ideas incorporated. This is one of the most interesting problems to solve in AI over the next twelve months, and we’re starting to see little signals of it in Claude Code and Codex (live artifacts).

17. The build vs. buy decision inside enterprises is at a very interesting moment. In most cases, I think it’s fully flipped to the point where build is the incumbent and buy has to make a case rather than vice versa.

18. If you’re not thinking about a company/brand brain, it’s a good time to start. Building tooling for the harness-first workforce to have access to all the systems and data it/they need is critical.

19. To that end, we’ll be talking more and more about agent experience (AX) and what it looks like to truly design agent-first. One of the most interesting questions in this realm is where AX drives you to make different decisions than UX/developer experience (DX) would. One obvious place is defaults. The rule of thumb we’ve set is to default these systems to human-readable output because the agents are much better at navigating complicated CLI flags and options.

20. Memory in chat/harnesses still doesn’t really work yet out of the box, and I’m not positive it ever will.

21. A lot of the productivity gains we’re seeing from AI are actually really smart people getting much better at using the models.

22. Evals still aren’t trustworthy. I frequently find that a model that has lower eval scores does better in real-world work than the maxed-out one.

23. I continue to be surprised that the new models seem to be worse writers, not better. I suspect there’s a mix of reasons for this, including the move to code/verifiability. But fundamentally I’m beginning to believe it’s because writing is fundamentally anti-consensus: if everyone is saying the same thing, it is no longer a good or interesting thing to say.

24. There’s a new model of work developing where you start by doing it with the AI, harden it into a skill, share the skill out to others, and then, when it’s sufficiently tested, figure out how to turn it into software. This seems like a pretty durable pattern.

25. In the past, the answer to “should we rewrite this from scratch?” was almost always no. That’s no longer the case.

26. While the labs train the models, much of the most interesting work is still happening at the point of execution where regular users find new capabilities by pointing them at their real work.

And a few from the original that I still think fundamentally belong:

27. We will continue to be surprised to find out that, without any specialized training, transformer-based models can solve many problems that specialized models were built to solve.

28. People are the bottleneck in enterprise AI adoption, and it will stay that way for longer than anyone expects.

29. If models didn’t progress from what we have today, we’d still have 10 years of runway to integrate their value into corporations and society.

30. When it comes to AI, don’t trust anyone who sounds too confident.

31. The value of using AI isn’t that it gives you great foresight into the future of how AI will evolve; it’s that it gives you an uncanny ability to sniff out everyone else’s BS.

32. AI is underhyped.

And finally:

33. This whole field is so fun. I can’t believe I get to do this for a living.

Editor’s note from Mike: We would love to hear any stream of consciousness thoughts about AI on all of your minds. Please feel free to build on the discussion in the comments or email me any takes at mike@brxnd.ai and we’ll incorporate into future editions of BRXND. Thanks, as always for reading.

Purchase Early Bird Ticket

An Inflection Point for ChatGPT Ads

There’s a bizarre bearish narrative around ChatGPT’s advertising ambitions making the rounds, lampooning the company’s internal projections of building a $100B ads business by suggesting the entirety of the “ad market for chatbots” is $5B. Here’s Eric Seufert last week:

An alternate information space seems to have developed around the chatbot advertising market that has led people to believe that OpenAI’s advertising product is an abject failure, ads in chatbots are fundamentally nonviable anyway, and the entire category will never generate a meaningful amount of revenue.

Like Eric, I find this perplexing for a number of reasons, most of all the fact that it’s near impossible to predict what form factor OpenAI’s core product will even take in twelve months. ChatGPT was a happy accident in the first place, and OpenAI has proved in the last week that it has no qualms about upending its entire user interface to pursue grander ambitions. Codex made a change to its core UI every single day last week.

While we talk about OpenAI speedrunning the Meta ads playbook, an under-appreciated aspect of OpenAI’s culture is how willing they are to make wholesale product changes with the confidence that they have a sticky enough value proposition that users will begrudgingly adapt. Few companies possess that level of product-market fit and audacity.

To this point, ChatGPT Ads has been a lean MVP iterating and shipping tentpole features at a frenetic rate. Three months ago, this was a closed beta selling a single creative placement at $60 CPMs! Thus, the only data point that has mattered so far about ChatGPT’s ad product so far has been the lack of user attrition and user engagement declines when ads are shown. In May, OpenAI reported “no impact on consumer trust rates” from ads being shown and has gone full damn the torpedoes in ramping up efforts to scale their ad business, hinting that the internal data backs up their public statements.

Now, things are about to get real. Last week, OpenAI emailed active clients announcing two marquee features—conversion optimization and app install attribution. Combined with the conversion API/server-side tracking, pixel and self-serve platform launch in May, OpenAI has now shipped the minimum viable product for direct response marketers to optimize for performance on the platform.

Now that conversions can be tracked, there’s going to be a natural obsession here to focus myopically on whether ChatGPT ads “work”, which is to say, whether they meet a brand’s ROAS goals in Q4 this year. Ignore that noise. Historically speaking, every platform that has reached humanity scale (i.e. a billion plus users) has managed to bolt on a highly performant ads business eventually. With an armada of Meta execs at the helm, OpenAI will as well.

The far more interesting conversation is around whether OpenAI can continue to scale a more aggressive ads product without flying too close to the sun on violating user trust. The design of the beta was intentionally minimalist, as if it was almost an advertisement apologizing to the user for being there. It’s likely that future ads will be far more intrusive – will they add enough value and serendipitous new product discovery to ChatGPT sessions that users will accept their presence? This still very much remains to be seen.

The other narrative worth watching closely is around when OpenAI begins experimenting with ads that are designed explicitly to persuade agents. To this point, all of ChatGPT’s (public-facing) ads vision assumes a human buyer and most of its ads pricing is moving to a cost per click model. Agents, of course, do not click. They don’t really view ads either. We’re in the earliest innings of “advertising to agents” taking shape as a concept and the most fundamental questions around what the business model will even look like are still wide open.

Per Juozas Kaziukenas, OpenAI is now experimenting with an ad format that will connect users with custom business agents rather than clicking out to external websites, perhaps a baby step towards thinking about how agent to agent flows will take shape.

The obvious tradeoff here is that OpenAI is asking advertisers to embrace a new level of disintermediation, unless marketing leaders can find a way to make their agents a true representation of their brand. To date, most agent workflows have focused on getting the basics right. The first order problem is nailing price, stock status and basic contextual questions consumers will ask about a product or service. But for marketers to hand more of the keys to agents, they need confidence that agents will operate as a natural extension of a company’s brand. This is a much murkier challenge and one that I suspect will become an incredibly hot area in enterprise AI adoption work.

In any event, the utilitarian promise of a successful ads product in ChatGPT far outweighs the risk posited by the cynics. Specifically:

1) A performant ads network in ChatGPT would be one of the most positive-sum, egalitarian drivers of growth— especially for small businesses both digital and analog– to come out of this era. If ChatGPT Ads does reach $100B of revenue by 2030, it will do so by driving hundreds of billions of dollars in growth across the economy. This success would also create vital competitive pressure on Google, Meta, and Amazon, who right now have enormous market power to raise CPMs.

2) The faster the ads business takes off, the bolder Sam Altman and OpenAI can be in their frontier bets, especially once OpenAI faces the whims of the public markets. I don’t think Sam took Ayahuasca one weekend and suddenly reversed his dogmatic contempt for ads. Someone showed him the financial model and cooler capitalist heads prevailed. The ability to generate digital rectangles out of thin air and sell them at 90%+ gross margins is truly one of the miracles of capitalism and it can help ensure the onramp to AI remains free for hundreds of millions of users.

3) The greatest career opportunity for up-and-coming hustlers in marketing will likely be to position themselves as closely as possible to the “ads in LLMs” zeitgeist, which will spawn both generational one-man businesses and venture-backed entities. It’s a hard time for many of the classic entry level positions in media and marketing– conventional media buyers, entry level performance marketers and all of the junior support functions brand-side and agencies alike are under intense pressure from AI. A net new platform rapidly gaining scale,, especially one that gen Z grew up with, would go a long way to filling the entry-level pipeline with marketing jobs.

To borrow inspiration from a famous presidential letter, ChatGPT’s success with ads is our industry’s success and I’m rooting hard for them.

Consumer AI’s Claude Code Moment

Across a lot of the more normie quotidian use cases, there’s a pretty defensible argument that the lived experience of using AI has been stuck in neutral over the past year. This is an interesting paradox– AI products can do so much more than they could a year ago, across both work and everyday personal use cases. Browser and computer use in particular unlocks fundamental changes in how people can interact with technology. In just a few months, computer use has gone from mystifyingly frustrating to near magical.

But it’s worth remembering how little of the population has experienced any of this wave in the grand scheme of things. Cross-referencing some data from ChatGPT for Work, Claude Code, Cursor, and Github, a reasonable estimate is that around 15-20M people are using coding scaffolds and advanced agent tools, many still within the boundaries of significant enterprise restrictions.

There are roughly one billion knowledge workers in the global workforce. At best, 2% are operating like most readers of this newsletter.

If your frame of reference for the “AI boom” of the last year is purely based on using mass market consumer products and trying to get Copilot to reference the Word document you’re working in, I can understand how you might be skeptical of what your AI-pilled friends are telling you! None of this is manifesting in your life!

Sam Altman does AI no favors here– he’s built a product that can automate the most mundane aspects of daily existence and magically free up more time to spend with loved ones and he cites an example of outsourcing parenting to AI as the utopia?

Beyond “better search engine”, the main consumer AI use case that has gained mass-scale traction is companion apps. Character AI has ~22M monthly active users, with average sessions in excess of 15 minutes. As a category, companion apps have around 50M monthly actives, with about 68% opening every day. It bears remembering that for how many people interact with AI, ChatGPT 4o was the pinnacle of the experience.

Much like adoption in the enterprise, consumer AI is no longer a raw intelligence or technology problem. In my first-ever piece for BRXND on rethinking agentic commerce, I wrote that “what’s been missing to date is a heavy dose of panache, audacity, and first-principles thinking in designing new front-end interfaces for AI.” This is still the first-order challenge across most consumer use cases.

It’s hard to say exactly what the catalyst will be, but I think we’re on the cusp of a massive step forward in consumer AI akin to what Claude Code and Codex brought to the enterprise. One interesting model for how to think about this is that in a year, the vast majority of knowledge work will happen in a harness and the world isn’t ready for what that will look like. What would it take for the same to be true of consumer AI use? I expect a couple of things to transpire slowly, then all at once.

1) The mobile harnesses will get much more user friendly. Codex Remote launched in May and already provides a model and harness combo that can run locally on an iPhone without taking up a gargantuan amount of memory. Voice mode has suddenly gotten really, really good. For now, OpenAI’s UX makes it hard for any non-Codex user to understand exactly what this means for their phone’s newfound capabilities but this is more of a marketing and design challenge. Thus, a really fun story in AI will be whether ChatGPT can launch their consumer super app before an open-source model provider builds a harness that is far more normie friendly.

Apple is in a fun position here. The company ostensibly “missed” the AI boom by not building or acquiring a foundation model and has seen its market cap rise by $2T (!) since the start of the year to reclaim the title of most valuable company in the world. If open source wins out, few companies are better positioned to win the race to $10T and Apple’s decision to sit on the sidelines and shrug off your insults about Siri will be viewed as an all-time corporate strategic masterclass. That said, you can feel something on the air that Apple’s new leadership has a rabbit in their hat that when produced, will usher in the next big step change in the industry.

2) Simultaneously, I think we’ll see a proliferation of niche consumer AI products that create very narrowly scoped moments of magic for users. Biographer, founded by USV partner Jared Hecht, is a perfect microcosm of this genre. The company leverages AI to help people tell the story of their lives, opening up what was once a highly expensive concierge service to the masses.

Many such businesses long to be built.

A few more things on my mind….

The team at Alephic created an outstanding video on Claude CoWork for Marketing, covering both the 101 of utilizing CoWork and more advanced concepts around context windows, governance and using subagents. The whole video is outstanding but per my section above, if you’ve never tried browser or computer use—or if you gave it a whirl four months ago and were exasperated by the clunkiness— watch the section of the video at 25:00 and give it a go again.

LinkedIn won main character of the internet status with a “Seems Like AI Slop” button that rolled out to users in the wild late last week. It’s been a very, very long time since I’ve seen a new feature launch from big tech that was as universally revered, in no small part due to the sublime irony of LinkedIn playing anti-slop crusader. Maybe nature really is healing.

Speaking of slop, I’m still perplexed by why Substack CEO Chris Best specifically used the term “Claudefishing” to describe the slop phenomena overtaking the internet. Obviously the more refined term for copy pasting raw LLM generated content is “Claude-dogging.”

Jokes aside, while Claude’s slop tells are a tad more pompous, slop is an egalitarian phenomena that is present in all models. Why would Best single out Claude specifically? My guess is that OpenAI and Substack are working on a larger partnership deal and Substack didn’t want the Pangram partnership to rock the boat any further there.

I’d expect Substack’s largest creators to start pushing for more features around how they get compensated by LLMs. Public writing on Substack is accessible by default, meaning external AI web-crawlers and third-party LLM developers can scrape open posts to train their models unless an author takes action.

Daydream, an AI shopping platform for fashion, is now selling its underlying search technology to brands. Founded by former Stitch Fix and Nordstrom executive Julie Borstein, Daydream was a fascinating canary in the coal mine for AI-powered shopping. The company raised $50M and launched a web search engine and app dedicated to providing a better experience to complex queries like “Sleeveless wedding guest dress for formal summer wedding”, a service the company is now selling B2B to apparel companies.

This is likely to be microcosmic of a much larger trend in and around AI where consumer concepts half-pivot into B2B as user acquisition efforts stall. An under-appreciated aspect of building in this era is that the cost of growth is significantly higher than it was in the Meta-Youtube-Google fueled arbitrage days of the 2010s, making it much harder to growth hack into a few million users. Power laws are also stronger now.

Ultimately, this sets us up for a world where there are a few mass-market consumer AI products and a longtail of weird esoteric niche concepts that definitionally are never meant to scale. I’m not sure there’s a whole lot of space for a VC-backed middle class of consumer AI apps.

ClickUp is hiring a CMO or “100x marketer” to lead a one person marketing department (+thousands of agents), at a salary of $500K to $1M. I sorta think of experiments like this in the vein of Bryan Johnson’s longevity quest. I wouldn’t want to be that guy but it’s good to have somebody out there pushing the frontier beyond its logical limit. In the same vein, I can’t fathom how you could possibly construct a brand marketing function this way but I guess we’ll find out.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

GET YOUR EARLY BIRD TICKETS

Paris, 1889.

The Exposition Universelle (world’s fair) opens on the hundredth anniversary of the Revolution, the Eiffel Tower newly unveiled, rises above it - the tallest structure on earth and a testament to engineering confidence. That January, the King of Sweden awards Henri Poincaré a prize for progress on the hardest problem in celestial mechanics: the three body problem, three bodies, each pulling on the others. (Hold that thought.) And in Clermont-Ferrand, two brothers take over a failing rubber factory and form a company called Michelin.

In retrospect, 1889 stands at the threshold of a century of management confidence: a belief that work could be broken into measurable units, organizations made legible - a world of future forecasting and annual plans. The Eiffel Tower expressed that confidence in iron while Poincaré’s work planted the seeds of warning from mathematics: even a system governed by known laws might refuse to yield a predictable future.

Working in modern corporate America in 2026, you can feel the chaos all around you. In my conversations with peers across industries the feeling is universal and the same themes emerge: management is acting poorly, channels are changing faster than companies can respond, customers are changing their behavior and everyone is producing (and being asked to consume) more work than ever before. The cohesion of work is evaporating - increasingly things don’t make sense.

Importantly, organizations, teams, and individuals have lost the ability to predict the future, and all of the scaffolding of “doing work” - annual planning, quarterly offsites, weekly status meetings - is increasingly feeling like theatre from an abandoned age.

Once again we are on the threshold of a new management era, ushering in… who knows yet? New ways of working with (and for?) AI. You can feel the doorway between the old world and the new world.

And perhaps no team feels this more viscerally than marketing teams - where AI is driving rapid change across three core areas: customers behaving in new ways, channels reinventing themselves and craft, the way work is made, is being completely rewritten. All three bodies are pulling on each other and the future dissolves into uncertainty.

Trying to predict the motion of three interconnected bodies is the definition of the three body problem that Poincaré worked on in 1889. He realized that once three bodies are pulling on each other, everything gets strange and prediction becomes impossible.

What does it feel like, in 2026, to lose the ability to predict the future? How do you coordinate work in an unstable environment? How do you navigate in times of chaos?

My theory is that the loss of prediction does not mean the loss of action. It means that the work becomes navigation.

Act 1: The Three Bodies in Motion: Customers, Channels & Craft

For marketing teams, the customer, the channel, and the craft are interconnected. The image of them as bodies in space exerting gravity is accurate: they exert force on each other. And all of them are changing at an accelerating pace in 2026.

Body #1: Customers

The old adage is that customer behavior changes slowly - that despite a technological revolution like the launch of the iPhone, it still took a decade for mobile usage to outstrip desktop. But I don’t buy it. This one feels different and the pace of change is astonishing.

ChatGPT and Google AI mode each have > 1B mau.

Pew 2026 says about four in ten U.S. adults use chatbots for information searching, and 38% of employed adults use them for work tasks.

AI is not just “information but now with natural language” - it’s about intelligence too cheap to meter.

Yes prompts are 5-20x longer than search queries (and getting longer). But the real shift is about delegation: customers delegating research and decision-making to LLM models. Where search touched the middle part of the funnel - useful for exploring a purchase decision and a small amount of discovery - AI is capable of recommending novel brands and categories to you, staying with you through your buying journey, and ingesting your preferences and constraints to get you to the final purchase decision.

On a recent family vacation to Vancouver, BC I had a “tom guide skill” that pulled local knowledge, insights, and recommendations all from my phone. Combining my personal context (how many kids, when we’re traveling, what we like and when low tide is). Without that skill I wouldn’t have seen this rad purple starfish at Willow Point Reef:

Vacation photos aside, the point is that the customer is becoming agentic. Not in the breathless buzzword sense, but in a practical way: a growing number of people now move through the world with a kind of cognitive sidecar. A model that reads before they read. A model that compares before they compare. A model that can scour, summarize and synthesize before they ever reach your site.

Quantifying this change is hard, especially because so much AI influence is either zero-click or dark-click. But we know adoption is growing quickly, and we can feel the shape of the customer request changing.

This matters because a brand used to be interpreted mostly by people. Now it’s interpreted by people plus the machine-shaped memory of the market.

The first body in motion - customers changing behavior.

Body #2: Channels

The environment that consumers are living in is changing too. As McLuhan says, the medium “shapes and controls the scale and form of human association and action.”

Search and social - intention and attention - the two biggest channels are both undergoing complete revolutions because of AI. The medium is changing.

Search & the information economy

Search is the obvious channel to change because of AI - but the second order consequences are still playing out and certainly not stable yet. AI search (AI Overviews and AI Mode) rewrites the implicit contract between searcher, search engine and brand.

We’ve moved from a list of links in ranking position to an AI model that talks about your brand - it can qualitatively recommend your brand (or warn against!)

Like a new predator being introduced into a stable ecosystem, AI search has disrupted the natural balance and order of things. 79% of top news sites block at least one bot used for AI training, while 71% block at least one live search or retrieval bot.

In the old web, almost nobody wanted to block search engines. Now publishers are putting up walls and making fine-grained decisions about training, retrieval, indexing, citation, and value exchange.

Google has teased a fully agentic future vision of search at I/O, but how much of this comes to pass if all the content is invisible to Google’s crawlers? The stable ecosystem of search is becoming disintermediated and it’s not clear how the story ends. Content licensing marketplaces? New forms of content monetization? New copyright laws?

Social is becoming AI spam

Social, once a place to connect with real people, then a place to browse algorithmic feeds, has now become an endless slop slot machine.

Social has changed less because of what the big platforms are doing, and more because of what they are NOT doing. Facebook and Pinterest have been slow and lazy in dealing with their AI spam problems.

But it is not just sloppy, it’s actual spam. Motivated actors are actively spamming these platforms for profit, producing synthetic media at scale to harvest attention, clicks, affiliate revenue, or platform payouts.

This is not a marginal problem. In certain categories, spam can feel like more than half the feed. More fake than real.

As the Guardian says “one man’s shrimp Jesus is another man’s side hustle”. And have you looked at LinkedIn / X / Substack lately? Pangram scores are being thrown around like red cards.

It’s slop as far as the eye can scroll.

(Worth noting that we are under-appreciating the flip side of this equation - explosion in creativity that is just around the corner as everyone has AI creative capabilities in their pocket. Knicks memes and animated pigeons.)

Two bodies in motion - customer behavior and channels both pulling on each other.

Body #3: The Craft

The third body is the way the marketing team itself works. The craft of work is changing.

It almost goes without saying that the world of work is changing, but we should not become numb to it. We are so early. While “AI” has been around for a few years now, we are really just getting started.

I think the advent of Claude Code and Codex was a bigger deal than the original ChatGPT 3.5 moment. AI usage has absolutely exploded because of these agentic tools. In May this year Cloudflare said it’s own internal usage of AI has increased by more than 600% in the last three months alone and Coinbase has seen token usage explode in just the last few weeks.

But the agentic revolution (Claude Code / Codex) outside of engineering teams is…. Maybe 2 months old at best??

Codex usage outside of engineering teams has just exploded, but only in May/June 2026:

If consumers have a cognitive sidecar with them, every knowledge worker now has a cognitive office with them, filled with personal staff: researchers, engineers, writers, designers.

For marketing teams in particular, 2026 has seen a Cambrian explosion in image and video capabilities, with SeedDance, OpenAI image models, Gemini, and so much more. Personally I’ve been enjoying generating, editing and producing video…. All in code? You can just ask an LLM to manipulate video files now. (see also hyperframes).

On the one hand the bar for creative output has never been higher, on the other hand you’re surrounded by slop cannons. Every knowledge worker is inundated by sloppy output and being asked to do more than ever - their old job, learn new AI tools and handle their coworkers 10x-ing their output.

So if the way work is being made feels new, different and strange - you are not alone. The two bottlenecks are now imagination and coordination.

All of this brings a sense of uncertainty. You can recreate workflows in AI, but the old workflows are out. Customers and channels are changing, so what new workflows do we need? How should the team be structured?

This is where the three bodies start pulling on each other and the future dissolves into uncertainty. If customers are changing and channels are changing, automating the old process is not obviously progress. You may simply be making yesterday’s marketing faster. Or worse, marketing to a customer that no longer exists in a channel that has changed.

So all three bodies are in motion: customers, channels and craft. The system has become impossible to predict.

What do you do about it?

Act 2: What kind of organization can act inside instability?

Organizations exist to solve coordination problems. Coase’s theory of the firm says markets are powerful, but using them is not free. You have to find people, negotiate, contract, monitor, remember, enforce, and decide. Firms exist because some work is easier to coordinate inside an organization than outside in the market.

AI changes the cost structure of coordination - individual throughput is exponentially increasing, but cheaper production is not the same thing as cheaper coordination! In fact cheaper production can make coordination harder - there’s more work to review, more mis-aligned output, more experiments, more variants than ever before.

More things that look done but are not actually decided. More things that look like progress but just place the burden on someone else to review and feedback.

The clock of coordination is the rhythm and rituals of sharing status (see: WTFS). Every organization runs on daily standups, weekly status reviews, monthly business reviews, quarterly board meetings and annual offsites.

But when the world changes the clocks change. Before the railroad, every town kept its own time - noon was whenever the sun said so. Then trains moved faster than town clocks, and we invented standard time so that everyone could stay in sync. And when the factory reorganized work itself, the first thing workers demanded was a new clock: back to 1889 - May Day was called in Paris in 1889, with workers demanding eight hours for work, eight for rest, eight for what we will.

New technology, new tempo of work, new clock. Every time.

If we accept the premise that your organization is going to be agentic - it’s going to be a mixture of agents and humans coordinating work there’s two ideas that become practical:

First, agents need context before they can be autonomous. Much like business analytics only works if you have a data lake, a shared context for all the data, maybe agents only work if you have a context layer. A shared, accessible layer of the necessary organizational context that is maintained and updated and agents can reference.

Second, human-made status dashboards are too slow for agents. The bottleneck is not production, but status. What needs working on, where is the project at? How do we adapt the status clock and tempo in an age of agents?

Railroads needed standard time. Maybe agentic companies need standard status.

We keep seeing these patterns at Alephic - the agentic layer only becomes useful after the context layer. Some examples from recent work:

Alephic Intelligence is a context layer: it is a centralized layer that ingests all of our organizational context: every meeting transcript, every slack message, every hubspot record, every code commit. This is useful for humans to query and reference but more valuable for agents to query and reference - now that Alephic Intelligence is built we’re just starting to layer on top automatic project status:

Amazon StarSearch is a project Alephic worked on with Amazon to import all Amazon reviews into a context layer. Once you have the context layer, it enables new kinds of marketing activity - like having Adam Driver read aloud a dutch oven review.

This is new kinds of marketing enabled by new technology, not automation of workflows.

Marketing Orchestration we’re working on a project with a Fortune50 brand to enable them to build a context-layer for their marketing team. A source of truth for all of the projects, status, actions, assets across hundreds of marketing team members and thousands of fast moving projects. We’re building the agentic layer on top of that to automatically move work around but you can’t do that until the context layer is built.

These examples all point to something practical for marketing leaders to invest in despite the uncertainty: build a context layer for your organization that is maintained, accessible to agents and leaves pheromone trails so agents (and humans) can operate on standard status time.

In my piece Of Termites & Tokens I argue that as the throughput of work increases you need to enable coordination through traces left in the environment.

Organizations have their own pheromone trails today: Slack threads, CRM fields, issue states, tags, transcripts, dashboards, source trails, approval markers, customer notes, analytics events, meeting summaries. In a low-throughput organization these are mostly residue. In an AI-accelerated organization they become the context layer that enables coordination via “standard status” (or token time as Justin McCarthy talked about with Noah).

If everyone can make more work, the company needs better traces. Better status. Better provenance. Better memory. Better ways of knowing what is alive, what is dead, what is approved, what is speculative, what is useful, what is merely plausible.

Otherwise the organization becomes a factory for its own confusion.

Of course, the interesting thing is not just ingesting the same context as everyone else (meeting transcripts, slack messages etc) but in ingesting the context specific to your business. For Amazon that’s StarSearch, the context layer of customer reviews. For another organization it might be a context layer of all the thought leadership produced globally by your organization and your competitors. For a biotech company it might be a context layer of all academic research, which lab produced it and who funded it.

AI of course makes this possible and easier than ever.

Act 3: Your Customers Are Lost Too.

There is another side to this. Your customers are lost too.

The sense of chaos and instability that we outlined above points to a feeling in the world: customers are as lost as you are. What do customers want when the world becomes hard to predict?

They want almanacs. Maybe.

The year is 1889, the same year that Poincaré won the prize for his work on the three body problem, a tire company called Michelin was founded:

“It all started in Clermont-Ferrand (a small French town) in 1889, when brothers Andre and Edouard Michelin founded their world-famous tyre company, fuelled by a grand vision for France’s automobile industry at a time when there were fewer than 3,000 cars in the country.

To help motorists develop their trips - thereby boosting car and tyre sales and in turn - the Michelin brothers produced a small red guide filled with handy information for travellers, such as maps, information on how to change a tyre, where to fill up with fuel, and for the traveller in search of respite from the adventures of the day.”

Before the Michelin Guide was a restaurant guide, it was a guide for early drivers. A tire company realized that if you wanted to sell more tires, you might first have to make the world more drivable. So the guide gave people maps, hotels, repair information, and eventually restaurants worth traveling for. They expanded the imagination for what driving could be.

Consumers are inundated by products that have AI features that barely work. And billboards that advertise AI without a shred of context.

Maybe the best brand marketing of the next few years will be about creating field guides for a disoriented customer. Almanacs. Maps. Buyers’ guides.

• What if a brand like Etsy built tools, education and almanacs on how to use ChatGPT for searching the vast Etsy library? (after all, Shopify’s catalog API is already allowing AI agents to search across all Shopify stores)

• What if a brand like Target or IKEA built guides for using AI in interior design? Not the cheap version of AI-generated rooms but rather something more mundane, real and useful. An almanac.

• What if a brand like Walmart built a guide for meal planning using AI? Educating consumers on what’s possible (recipe discovery!) and what’s hard (AI-recipes) in 2026 and how and where to use the AI tools to help.

Of course AI is amazing at shopping, interior design and meal planning. It is also terrible at all these things. For a brand to pull this off there has to be an honesty to the degree to which AI can be used in practical ways for practical tasks, and a recognition of the limitations and flaws of the current models (where those flaws and limitations are moving very fast!). Maybe marketing’s task is not just GEO (which is just SEO in a trench coat) but it’s about building your own evals and benchmarks for how well AI helps consumers in your category.

Before the brand marketing team at Walmart can build a consumer guide they need an internal WalmartBench eval framework for how good the various models are at finding products, searching their catalog and guiding customers in useful ways through their purchase journey for the categories that matter for your brand.

Maybe your customer needs an almanac.

This is, I think, the marketing opportunity inside the three body problem. If customers are changing, channels are changing, and craft is changing, then consumers are hungry for orientation.

How does the story end? The mania of new machines.

Let’s return to Paris in 1889, the Exposition Universelle but this time at night. It’s the first world’s fair to stay open at night because of the invention of electricity. Thomas Edison was a big feature at the fair and illuminated fountains with colored electric light captured the imagination of the people with a spectacle never seen before.

“In the basin, figures play in agitated poses, with tritons spouting torrents of water. The whole ensemble is highly decorative. Manufactured using new processes hitherto unknown in France, this fountain, illuminated each evening by electric light, shoots jets of colored water to great heights, which radiate into a dusting of alternating red, violet, blue, green, or yellow. The sight is truly miraculous.”

Meanwhile, just a few years before in 1882, Scientific American Supplement published an essay on electro-mania where W. Mattieu Williams argued for the separation from hype and reality around electricity - highlighting the mania of electricity making dead frogs move and people believing in re-animation.

Fast forward to 2026 and the essay by Nikhil Suresh: AI Mania Is Eviscerating Global Decision-Making feels eerily similar. I agree with Nikhil’s headline, that AI-mania is eviscerating decision making, but not the conclusion that all is lost. Chaos is all around, but we must learn to act regardless. New technologies always come with a breathless mania - our task (perhaps the uniquely human task?) is to separate the spectacle from the sublime. To discern between being tasked with animating dead frogs and illuminating dancing fountains.

The answer to a world that can no longer be predicted is not better prediction. It is better orientation. Prediction died in 1889 but navigation didn’t.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

A brand used to be what you said about it. Now it’s about what a reviewer, algo, creator, or agent says. It’s not that the brand doesn’t have any say, but it’s likely out-voted.

My LinkedIn feed has lots of really interesting conversations about the problem, but not so much about what to do about it. How does a brand build the meaning it wants to build in this environment? Which brands are built to thrive? What does the discipline of brand positioning and comms planning even look like anymore? It feels like folklore to me, bits of stories passed around, with people and machines adding their own bits to the stories as they go.

This newsletter has two things in it: The first is an experiment to see whether we can test whether a brand is built well and whether it’s being “retold” in a way that’s useful to the brand. The second is an essay about a possible path forward for the people who are in charge of brands.

An experiment: The Retellability Test

How do we evaluate whether a given brand is able to build meaning in this environment? A starting hypothesis: We need to look at two things. First, whether the brand is ready to be retold or carried. Second, whether it’s actually being retold accurately and favorably.

This test scores 15 variables, all pulled from public evidence. None of these brands asked to be in it. Together, the variables make up two key measures:

1. Brand Readiness (9 variables): Whether the brand is built to be retold by other people and machines (whether it has distinctive assets, a world worth repeating, a record clear enough to attach things to)

2. Brand Rendering (6 variables): Whether it actually is being retold by them (whether they’re recommending it, discussing it, and writing about it), and whether what shows up is accurate and favorable

Consider this a starting point. The whole process will undoubtedly get better as we go on.

Here’s an example of how it works: In the sunscreen category, Coppertone scores 51 on Brand Readiness and 47 on Brand Rendering, which is decidedly middle of the pack among the brands we measured.

The diagnostic found that Coppertone has the deepest heritage and the most complete public record in the category. Eighty years of history, a photo archive, twelve years of Reddit threads. But almost none of it converts. The brand is not recommended or retold in any meaningful way. When people and machines talk about sunscreen, Coppertone is mostly absent or, every now and again, it’s the affordable one that smells good.

Maybe counterintuitively, this implies that a deep brand heritage does not necessarily mean it will be retold. One hypothesis is that Coppertone’s brand feels monolithic and finished. It’s not something that either people or machines want to share or recommend. Leading it to be cited, but not retold.

The new folklore

In June 2026, the CMO of one of the world’s largest advertisers, Unilever, delivered a keynote at Cannes and said his company had “stopped asking what our brands want to say” and started asking “who would care enough to carry it.” Here’s his talk.

That one sentence sums up where the marketing world finds itself today as well as anything I’ve read. “A brand is what people say when you’re not in the room,” the old saying goes. It’s always been true. This is folklore. A story passed from mouth to mouth, changed a little by each teller, and (importantly) owned by no one. Cinderella, for example, has countless variations across time and cultures, but we all recognize it when we hear it. Brands work the same way.

Historically, brands would heavily contribute to their own folklore by authoring stories and sending them into the world. But now, the balance has tipped. We don’t control our brand’s stories anywhere near as much as others do. Today, a brand is far less what it says about itself, and far more about what a wide ecosystem of third parties, both human and non-human, can find, trust, rank, recommend, and sell.

Lots of people are already talking about this issue, not only Unilever. Almost all of the conversation, though, has been about identifying and naming the problem. It goes something like this: A brand is a shortcut for meaning in our minds, but how do we keep that meaning coherent and effective when we aren’t the ones making it any more?

We used to be able to control a lot of the brand’s ecosystem. We could pay for a thirty-second spot, better shelf placement, and a mention in a top-tier magazine. We staged the launch and approved all the outputs. We held that control-everything stance even as production got cheaper and sped up. The discipline for putting guardrails on all of those brand touchpoints was called comms planning, or engagement planning.

Now that world has changed, and we cannot control the great majority of our brand’s meaning-making. A reviewer writes a pointed post on Reddit or Substack. An AI agent turns it into three sentences a shopper reads instead of your site. A creator turns your brand into a routine, or a joke, or just the latest “can’t miss” hook. A fleet of them spams those hooks at their followers, racing each other to the best ROI. A retailer turns the brand into shelf position or sponsored search. A marketplace turns it into a thumbnail, a review count, and a rank.

This is a real problem, and I’m here for it. But the more interesting question is what we do next?

The reflex is to make more stuff. More is more. Something will hit, and we’ll just amplify that. If done well, that’s not wrong, but there is fallout. Meanings multiply and contradict. The brand turns to mush.

“Because we do mass volume, there’s no way this can’t work.”

— some dude in a sponsored post on my FYP, selling access to a TikTok creator pool, July 2026

How do we build brands to be better equipped for this environment? A well-equipped brand, when it is “carried” (to use Unilever’s word), or “retold” (to use mine), gets stronger through retelling. And it does so in a consistent way, with lots of people, over time.

I think the answer lies in better upstream conditions or incentive design.

The old way is executional control, where we carefully shape every output, one at a time. But we cannot possibly approve every Reddit thread, every comparison article, every creator riff, every AI answer. Tighter guidelines just can’t scale across all of that. That lever is disappearing.

Instead, imagine a discipline that lives upstream of the go-to-market plan. This discipline is responsible for designing the conditions and incentives that make good outputs more likely: The codes, claims, participation hooks, incentives, and feedback loops that let a myriad of outside actors, both human and non-human, produce something recognizable without approving every single thing. This discipline probably shouldn’t even be called comms planning anymore.

We are still after the same end goal, namely a strong brand that is able to build those consistent memory structures that drive buying decisions. But we are building the conditions for it, not building the touchpoints themselves.

Letting go of control is difficult. It means letting others play around with the brand in their own way, for their own reasons, and using their own words.

What’s more, it’s difficult to design the conditions and incentives to make good outcomes more likely, because the thinking has to happen up front. Get the inputs wrong, and the ecosystem compounds the error at a scale that the brand police cannot possibly catch up to.

Brand meaning will be made whether the brand participates or not. If it doesn’t participate, the brand doesn’t fade to nothing. It just slow-drifts until it means something else entirely.

I have theories about how to do it, and have started using them for clients. But no one has the exact right answer yet. We all need to explore and experiment more. My first public experiment is the Retellability Test. It tries to measure whether a given brand is built to be retold (Brand Readiness), and whether others are actually retelling it accurately and favorably (Brand Rendering).

The Unilever CMO says brands should be fires, and creators are the people who carry them. It’s a good metaphor, but not a new one. Folklore has always worked this way. The best stories have always survived because they got retold, not because anyone owned them.

So the part I care about, and the part not enough people are talking about, is how you build a fire worth retelling in the first place. What’s that fire made of? Why do some fires catch but most go out? Which brands are doing it well already?

We should be able to build brands strong enough to stay themselves, no matter who’s telling their story. And I believe, if we build them well, the retelling will make them stronger.

If you have any questions, please be in touch. As always, thanks for reading.

— Noah

GET YOUR EARLY BIRD TICKETS

The Enduring Power of Slop

The terminally online are having a viscerally anti-slop moment. Gradually, then suddenly, the tech elite have decided they’ve had enough of sacrificing the written word at the syntax abattoir of “it’s not delivery, it’s DiGiorno.” Substack in particular threw down the gauntlet this week, partnering with Pangram to launch an AI detection feature across the platform.

For most of us working in and around marketing and AI, this feels like the proverbial “nature is healing” moment. But to most of the world, we’re just a bunch of whiny wordcels. Based on sheer engagement metrics, the people love empty calories masquerading as a healthy meal.

Outside the filter bubble, the proliferation of slop continues at an astounding pace. >40% of longfrom LinkedIn posts are now AI generated and shameless copy paste jobs regularly go viral on the platform. On X, one of the most shared pieces around the World Cup was this AI-generated essay hyping up the US. Soccer team. Even as someone who has Pangram wired into my consciousness, I read it and a small part of me wanted to run through a brick wall. Matt Freese punished me accordingly.

Slop, of course, is not an entirely new phenomenon and it long predates LLMs. NCIS…..ah whatever the hell city they’re in now… was the top watched drama of the 2010s. We have endured ten Fast and the Furious movies, not even counting the bizarrely excellent sloptacular known as Hobbes & Shaw. Plenty of human written content, especially when forced to take the form factor of SEO arbitrage, was just a more simplistic slop varietal.

Slop endures because it is ultimately a demand side problem. LLMs don’t even create our inherent demand for slop. They simply satiate it. Said another way, slop has product-market fit, at least when it comes to influencing humans. Paradoxically, I agree wholeheartedly with Noah’s take from last year that slop will be far less effective at influencing agents long-term than it is at hypnotizing people.

But while a vocal minority seems to be reaching their slop breaking point, why does AI-generated text maintain such a stranglehold over its consumers?

Slop wins because it conforms to neat narrative arcs that fulfill our desires for cohesive stories: At crime scenes, children make the best witnesses because they simply report back what they saw rather than editorializing into a natural story arc. The adult brain is hardwired to eliminate cognitive dissonance by effectively hallucinating events that didn’t happen to make our memories fit into a tidy narrative.

Does that sound like any of the AI tools you know? In many ways, a tragic flaw of LLMs is that they are even more allergic to cognitive dissonance than people are, hence the hallucination problem being such a tough one to squash completely. Moreover, any time there is even mild tension or paradoxical sentiment in an idea, models go out of their way to call it out ( “that’s the real sticking point”) and try to crush it when the more intellectually honest tactic would be to simply let the tensions be.

Insufferable vernacular tics aside, this makes AI writing more “comfortable” to read, especially in the passive consumption mindset that, well let’s face it, most people are in most of the time.

In many ways, the most interesting thing to talk about when it comes to slop is why it seems to those of us who spend the most time with these models that each new model, despite its increases in intelligence, seems to be more slop-y, not less.

Here’s Noah’s observation from earlier this month:

One of the things I’ve found most confounding over the last year is that as these models get better, at the same time, I feel like they’re getting worse at writing. Or maybe “worse” isn’t quite the right word, but I’m not seeing any end in sight for the slop.

I suspect this comes down to a confluence of circumstances: the nature of the model is to average things out, reinforcement learning with human feedback (RLHF) moves the models towards being a “helpful assistant,” and the move to coding and reinforcement learning with verifiable rewards (RLVR) means the models are further steered towards a model of the world with conclusive truths.

But my guess on the biggest challenge is that good writing is fundamentally anti-consensus, and I’m not sure that’s a problem the models can really solve. As soon as the models all agree on a turn of phrase or style, it becomes “slop” and can’t really be used by humans anymore. As opposed to code, where (for the most part) a good answer is a good answer (you can verify the code works), the quality of writing is subject to other writing that exists around it.

At this point, I’ve abandoned using AI to generate any copy at all for public-facing work. Perhaps there’s some dogmatic pearl clutching here and I’ll re-evaluate with each new model release but for now, the trend of each subsequent model doubling down on a slop-y vernacular identity (and increasing the conviction with which it does so) is accelerating.

With heavy training on my voice, I used to work with ChatGPT and Claude for writing transition paragraphs, flagging and eliminating weasel words, and pushing me more towards brevity. I always found LLMs made my prose too detached and omniscient but older models felt more malleable. I now feel mostly powerless to stop the models from injecting stylistic touches that simply do not sound like me.

Ultimately, I suspect that we’ll lose more of the internet to slop simply because in more and more instances, slop will get the job at hand done. But as with any good or service, there will always be a market for artisanal prose, a market that will grow as agents seek out higher quality content and increasingly esoteric expertise. Thus, the inherent value of writing that requires real learned experience and imbuing contrarian thought will also go up. On balance, this is a better era to be a writer than the pre-LLM times.

But for all of its captivating power, AI-generated text will always have one fundamental flaw. Even if it satisfies the reader, the writer is a cognitive passenger princess in their own life.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

GET YOUR EARLY BIRD TICKETS

When I originally announced BRXND in 2022, it was because I felt there was a better, more important conversation to be had in marketing about AI. At that time, it felt like all anyone wanted to discuss in the industry was the possibility of AI-generated television ads, and that felt (and still feels) like such a tiny piece of a much larger puzzle.

Fast forward to 2026, and I feel a bit of deja vu. While the industry has thankfully come to understand that the opportunity is bigger than just AI commercials, there’s another massive underestimation happening, but this time it’s around consumer adoption and what it means for brands. While most of the industry seems obsessed with getting on Reddit, my feeling is that this is the biggest shift in the way consumers consider, research, and buy products since the birth of ecommerce thirty years ago.

To that end, this year’s conference will build on our past focus of helping to highlight how the world’s best marketers are integrating AI into their organizations and work while also spending more time looking outward to consider what the future of brand building looks like in a world where preference is heavily mediated by large language models. To get things started, I’m incredibly excited to announce the initial set of speakers and sponsors. We’ll have more coming out over the coming weeks and months, but a huge thank you to those who are helping to bring this year’s event to life.

Initial speakers:

• Jo Shoesmith, Chief Creative Officer, Amazon

• Mark D’Arcy, Corporate VP and Global Creative Director, Microsoft AI

• Mikkel Holm, Chief AI and Innovation Officer, Bark

• Aaron Dignan, VP of Agentic Products, Rocket Money

• Mike Houston, Head of Creative AI, Amazon

• Tom Critchlow, Forward Deployed Engineer, Alephic & Founder, AI Search Leaders

• Justin Inman, Founder & CEO, Emberos

While the narrative has evolved as LLMs increasingly mediate consumer preferences, the core tenets of BRXND are the same as when I put on the first show in 2022.

Doers > Speculators: We want BRXND to be about doing and learning, not talking and speculating. We are too early in this journey, and the technology is too different from what we’ve seen before to get stuck speculating about its future impact.

Curiosity > Overconfidence: We embrace the unintuitive nature of AI and its future implications. We are all learners in this new era of technology.

Conversations > Panels: Whenever possible, it’s better to have just two people on stage talking to one another rather than a whole set of speakers competing for time and attention. Go deeper with fewer people so that the audience can learn from their thinking and experience more than just their opinions.

Show > Tell: It’s always better to show people something than to tell them about it. While other conferences try to urge people away from showing specific products, we think it’s most useful for people to see what’s real and possible.

Process > Perfection: Take your audience on a journey, even if it’s something that didn’t work out. This is an audience of doers and experimenters, and understanding the journey is often more important than just seeing the outcome.

This is just the beginning. We’ll be announcing additional speakers in the coming editions of the BRXND Dispatch as we build towards an incredible day on November 5 at the Times Center

Final reminder: early-bird ticket pricing is available through the end of the week. To add your name to the waitlist, please visit our website. Thanks for all of the support, and I look forward to seeing you in November.

GET YOUR EARLY BIRD TICKETS

BRXND NYC is coming up on 11/5, and early bird tickets are currently $749. Get yours early, we will sell out.

GET YOUR EARLY BIRD TICKETS

So while clearly token leaderboards don’t represent the long-term answer, I want to offer a defense of the practice of tokenmaxxing (and token maximalism generally), as by far the best approach to learning how to get the most out of these models. The thing I think most of these articles miss is that when it comes to driving change inside organizations (and society), people, not technology, are the bottleneck. If you fail to address that, then nothing else really matters.

One of the questions I get most often from customers—along with friends and family—is “what can I do to catch up in AI?” My answer for the last four years has basically been the same: get yourself the most expensive subscription you can and try to use as much of it as possible. It used to be ChatGPT; now it’s Codex and Claude Code/Cowork, but the advice stays the same. Specifically, I suggest you use it so much that you max out your token budget on their biggest plans.

To give a sense of just how many tokens that is, here’s a recent SemiAnalysis estimate on what you get out of both companies’ most expensive plans (the “max possible spend” number is their best guess for the value you can get out of the plan if you were to use it to its fully allowable extent):

If you’re really trying to wrap your head around AI, nothing replaces the hands-on keyboard experience. At the very first BRXND in 2023, I talked about building AI fingerspitzgefühl (fingertip feeling), and despite the years in between, I still think it’s the most critical piece of the puzzle, and I credit quite a bit of my own understanding and success with AI to just having been lucky enough to be able to put in a thousand-plus hours in 2022.

The reason nothing replaces hands-on also hits on a big misconception about AI: that you can just hand things over to it. While Fable and Sol get us closer than we’ve been before, for the most part, when you watch people who get the most out of these systems, you see them subtly steering them to success. Oftentimes it’s such small adjustments that even they don’t notice. It can be the way they ask the question or when they choose to interrupt, but the outcome is that by making one- or two-degree adjustments, they are experts at keeping the harness on track. The most insidious AI failure case isn’t it writing bad code or a bad doc, it’s fundamentally doing the wrong work.

Which brings me back to “tokenmaxxing,” the pejorative term people have come to use for companies that encourage their employees to use AI by any means necessary and often track them in leaderboards. Here’s my general view on the practice:

1. When people talk about tokenmaxxing, they’re mostly talking about a) the token leaderboards and b) a few reports from companies like Uber of massive AI spend.

2. I obviously don’t think leaderboards are the best long-term solution, but I’m not sure they’re the worst short-term solution (more on that in a bit).

3. While there are a few stories of massively inflated AI spending, I’d venture to guess that the vast majority (90%+) of the Fortune 500 should be spending far more on tokens than they do today.

4. When we hear about these companies spending real money on tokens, it’s happening via harnesses like Claude Code/Cowork and Codex, not ChatGPT, Claude, or Copilot. More importantly, that inflated spend mostly isn’t coming from tools built with AI, partly because those tools, by their nature, are much easier to optimize spend around.

5. The vast majority of the Fortune 500 doesn’t give their employee base access to platforms like Codex and Claude Code/Cowork.

6. Most people who work in large companies don’t want to change the way they work.

That last point matters a lot. If you’re trying to turn a massive organization, you are usually better off over-rotating than under-rotating. This is also why, in the short term, a blunt instrument like a leaderboard might actually work in motivating people to dive in. Obviously, in the long term, you want to see value from these tokens, but remember that the average Fortune 500 employee is currently using significantly fewer tokens than they should. Participation before optimization.

The conclusion of all this seems very clear to me: for a small group of technology companies, getting past tokenmaxxing really matters. They’ve convinced a large enough swath of employees to integrate these tools into their lives that they’re now ready to graduate to thinking about the return on those tokens. But for the other estimated 90-95% of large enterprises, the greater concern is that they’re currently token minimalists in a world where the return on knowledge and capability can be exponential. In part, this is a function of point number five, as most of these companies still haven’t opened access to the very best AI tooling (and I’m not quite sure what the path to that looks like). Like all things AI, the biggest mistake pundits make is forgetting that people are the bottleneck. This transition, despite the breakneck speed, will still take a long time to fully sweep through the corporate world, and we should all be wary of claims that we’ve entered a new phase.

Since you’ve made it to the bottom, one more reminder that BRXND NYC is coming up on 11/5, and early bird tickets are currently $749. Get yours early, we will sell out.

GET YOUR EARLY BIRD TICKETS

GET YOUR EARLY BIRD TICKETS

The GEO Leaderboard

The modern marketing leader’s inbox is a dystopian postmodern art form. If there’s any inkling that you might influence a software purchase, you live perennially awash in a deluge of pitches for identical-sounding software. Right now, the hottest technology is all about how brands show up in AI search, which, from here on out, I’ll describe using the generative engine optimization (GEO) lexicon.

Amidst the litany of market maps, VC graphics, and customer review boards covering the 50+ vendors that have appeared on this scene in the last two years, there was one question that I had not yet seen answered: Which company that is selling me the ability to be present when users are querying LLMs is best themselves at showing up when marketers search for GEO solutions?

Or said more simply, which GEO company is the best at GEO?

To answer that, I grabbed a couple of Cometeer capsules, fired up a DataforSEO account (thanks to an assist/suggestion from Zero.xyz), and spent an afternoon vibe-coding the resource I wished to see in the world.

What emerged is a simple leaderboard that ranks the top 10 companies based on how prominently they appear across 175 prompts that marketers would use to search for a GEO solution. The link to the live webpage is here:

In addition to the overall table ranking, for each company, you can drill down to see where a GEO vendor is most visible across model, prompt archetype, and specific queries:

I encourage you to explore the leaderboard, dive deep into the methodology and peruse the full set of prompts the tool tracks. Please send me any thoughts on what stands out or what I got wrong! I intend for this to be a fluid project that I’ll re-run monthly and add additional companies to.

To me, what’s far more interesting than the tool itself is what the process of building it revealed about both the nature of AI search and where much of the value in vibe coding experiments like this ultimately accrues.

Learnings and Observations

1. The high-level results were fundamentally pretty unsurprising… which is kind of the point! The companies with the strongest “brand” footprint in GEO rose to the top of the list.

Every day, AI search becomes less of a convoluted morass of grey hat arbitrage schemes. It’s much, much more akin to serving as a mirror for brand marketing!

As the models mature and get more sophisticated about how they access real-time web data, the hit rate of spamming Reddit, blasting listicles with your product at the top and indulging in programmatic content schemes will increasingly be reduced. In its place is just good consistent work across building credibility with top media, telling a clear and differentiated brand story and building clear owned content lanes and tentpoles.

2. LLMs have a frenetic, token-thirsty desire to build incredibly complex projects that test the boundaries of their capabilities. For this project, Claude Code (running Opus 4.8) was eager to build out a full Next.JS app with all kinds of interactive components that would effectively have constituted shipping the minimum viable product for a full-on GEO platform. It took a lot of heavy steering to dial this down to a basic output that could ship in hours, not days.

At one point in the session, I dropped a, “my brother in Christ, I’m not trying to fully recreate the ten companies on this list.” I’m simply trying to build a simple HTML tool to help marketers better understand which GEO companies are best at GEO!”

Even still, I ended up building an MVP with far too much wiring that Noah wisely suggested knocking down to the final webpages that you see on brxnd.ai.

3. At the same time, I pretty easily could have gone a step further and vibed up an application that looks a lot prettier and recreated something cosmetically close to the dashboards of GEO companies. Claude was begging me to do it!

This is not at all a dunk on the technology these companies have built—all of the serious executives I’ve spoken to in GEO are the first to tell you that basic prompt and citation tracking is a commodity. Albeit a commodity that has had enormous CMO-level demand to date.

The far more difficult and important question is understanding the sentiment and nuance of how LLMs perceive your brand and ultimately building agents and workflows that shape that perception and destiny. This is where we’ll start to see meaningfully different companies emerge in the space.

Profound is all-in on helping marketing leaders build agents, championing the concept of a marketing engineer. Evertune is building an ads manager to closely align GEO with the potentially far more lucrative whitespace of buying media in ChatGPT. Bluefish is now helping brands build agent-optimized catalogs that align SKUs to semantic brand stories, a vital GEO concept that is still in its early innings.

Whether GEO companies grow into their valuations depends on whether they can cross this chasm from analytics software to truly running end-to-end workflows where brands take action. More on what all this means for brands in a “GEO Maturity Matrix” post to come.

4. In the vein of Noah’s software for one experiments, it’s not that hard to see how an enterprising CMO (or more realistically, an AI-pilled member of their team) might decide it’s worth spending an afternoon vibe-coding a little project like this to make the best possible decision on which software to buy. Bain Capital has PE associates vibe-coding legacy SaaS replicas to understand whether their acquisition targets have true moats.

If a marketing leader took up this mantle and ended up in the same loop I did with Claude, they might just find themselves tempted to “build the thing” around their bespoke use case rather than modifying off-the-shelf AI software. Why pay tens of thousands of dollars per year for something you can build that is purpose-created for the brand metrics you want to track? More than most categories, AI search requires technology that touches a brand’s “unique judgment, workflow, brand voice and edge.” There’s a real build case here.

5. Before AI, the point of a great strategy doc like the famous Amazon six-pager was that it forced the person writing it to deeply grapple with an idea, identify fundamental flaws, and ultimately, sharpen the thinking around a proposed product launch or strategy. Driving executive and team alignment was a secondary benefit—the bulk of the value came from the agonizing clarity that only comes with deeply thinking through a problem.

At its best, vibe coding is just another means of pressure testing ideas and identifying fallacies in your thinking, only the medium is an artifact rather than a document. The challenge here, of course, is that LLMs seek to strip you of agency at every turn. They want to quickly make decisions, ship code, and yada yada over any paradoxes in a project. I consider myself an ultra-amateur vibe coder, which is to say, I still hand over way too much agency to LLMs when I work on things like this. But with each rep, I find myself slowly imbuing a little more “me” into my AI-assisted work.

When anyone can brainlessly build, doing the work is the point.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike

GET YOUR EARLY BIRD TICKETS

The Parallel Open Web

The argument for the obsolescence of the open internet we know and sometimes love is ultimately a pretty simple one.

When the largest technology companies in the world have a vested interest in protecting access to proprietary data, the web is destined to become a closed entity of fiefdoms that feed the AI beast. In a world where context provided to an LLM is king and distribution is the rate-limiting resource, the die has ostensibly been cast.

In response to the rational economic incentives AI creates, all of the best content providers will move behind paywalls to directly monetize their audiences and exert whatever leverage they can with LLMs for licensing deals. More than ever, information simply does not want to be free. It longs to be expensive and scarce to accrue maximum value to those who have it.

Coming out of I/O, it’s pretty clear that Google is fully at peace with ending any last vestiges of free content depending on search traffic. Ten blue links that direct users to external websites have served their purpose and in Google’s eyes, are no longer the best form factor for efficiently satisfying a user’s query and maximizing their own ARPU. The malthusian trap is worth the risk. Google does this matter of factly with malice towards none but with charity towards none as well.

But ironically, the thesis for the endurance of the open web is similar to the case for its demise, with one key twist. As agents come to massively outnumber humans online, there will be a Cambrian explosion of new demand for readily accessible content. This force will be stronger than the pull towards walled off centralization.

This idea is best encapsulated by Parallel Web Systems (and a few notable competitors such as Exa), founded by former Twitter CEO Parag Agrawal and valued at $2B. Parallel’s mission is simple– build the search and retrieval infrastructure that lets AI agents read the open web at scale.

As a result, Parallel has taken up the mantle as the champion of the “open web”, a term that has historically been co-opted by programmatic vendors to sell a chumbox of ad products from a bygone era that have rendered many media properties utterly impossible to consume. In Parallel’s recasting, the open web is “modern humanity’s living memory…and the ecosystem that made modern AI possible in the first place.”

Parallel’s thesis is effectively that for AI to fulfill its grand promise, the spirit of the open web must endure and find a new economic model to support openly indexed content that lives outside walled gardens. Without it, agents will become stale commodities.

Last month, Parallel launched a new product called Index, a crack at building a new economic layer for open web content. The platform shows content owners how agents discover, cite, and rely on their work and compensates them based on each source’s contribution to the agent’s output. It’s a manifestation of an idea Ben Thompson was all over a year ago when he argued that micropayments— a concept that fundamentally never made sense on a web built for human attention– could finally be viable on the agentic web.

Patrick Collison also has a good take:

What I’d emphasize is that while Index is nominally a publisher-side product, what it really ships is a new framework for influence. For a brand, economic value now increasingly accrues to your position in an agent’s reasoning at the moment of inference, not to a human’s attention at the moment of blog post read, ad impression or PDP view.

The end of the commodity internet

When it comes to which entities influence agents, it’s important to remember that the metric of 68% of searches now being “zero-click” disproportionately obliterates a specific segment of the open web ecosystem. SEO-dependent publications– which largely only existed in the first place to exploit keyword arbitrage– have been wrecked. Entities that publish timely news, analysis, entertainment or bizarre esoteria that serves real-time retrieval have been far more resilient than reported.

This last point is the most interesting to me because it lays out the path for the “open web” to once again return to its free-wheeling roots.

When monetizing web sites required reaching a critical scale of human eyeballs to sell ads to, you had to build within commoditized patterns that could both rank for SEO and serve necessary ads infrastructure. For 25+ years, the pendulum of web architecture has swung towards standardization, especially for commercial and marketing oriented builds.

If the business model gets rebuilt around serving AI at the point of inference, more of the “standardization” can exist at the invisible infrastructure level. Certain protocols will need to be in place for agents to retrieve information but effectively there’s a path for the open web to become a subliminal layer targeted mostly by agents on top of far bolder, wackier web interfaces built for humans.

I’m sympathetic to this case for open web renaissance for a few reasons:

- The arc of digital history generally bends towards content that can be accessed for free.

- There’s just a hell of a lot of pent up creative energy to build weird stuff right now with all of the new capabilities that AI unlocks.

- Simultaneously, there’s a vibe right now where a lot of people are just building for the pure love of the game, happy to worry about business viability later. The early blogosphere was crafted on a similar ethos.

- There’s massive uncertainty around what kind of information agents will ultimately gravitate towards as the models become more sophisticated. When nobody knows what the hell is going to work, the best strategy is for content creators and brands alike to maximize shots on goal.

Agents are weird…and perhaps the web will once again be too

If there’s a critical flaw in what Parallel is trying to build with Index, it’s that micropayments have also historically failed as a web business model precisely because they ascribe a rationality to the media business model that doesn’t exist. If media was a pure utilitarian purchase, most publishers would be extinct.

Quality media is largely sustained by people who buy the idea and self-actualization of being a New York Times, Atlantic or Stratechery reader more than the utility of actually reading. That is to say, they buy being perceived as erudite by their peers and pay hundreds of dollars for coffee table accessories. They buy the brand.

One popular theory of the future is that agents won’t be so sentimental. Agents, so the conventional logic goes, will rationally seek out the most relevant information to complete a given task and will pay fair market value for it, agnostic of “brand.” For media, that fair market value often doesn’t sustain the cost to create content.

Truth is, at this juncture, we have little idea how agents will ultimately hold a notion of brand preference. Hell, there may well be agents for maximally cost conscious consumers that render brand as more of a tax than a buying signal.

But a few things are likely to be true that point to agents building true brand associations:

- Agents will be increasingly trained to look for heuristic shortcuts that don’t swallow as much compute

- Agents will take on more quirky anthropomorphic tendencies as models get more sophisticated

- Influencing agents will become one of the most lucrative problems in tech. It will be far less straightforward that simply optimizing catalogs and stacking Reddit karma.

Taken together, it’s hard to understand how the increasingly complex context needs of agents will possibly be served if the web devolves into a morass of walled gardens, licensing deals and content marketplaces.

Perhaps it’s as simple as this— all signs point to the agentic web swinging us back towards things getting really weird again. And it’s the weird early 2000’s web, where people wrote entire blogs from the perspective of ficus trees, that so many of us long to see return.

If you have any questions, please be in touch. As always, thanks for reading.

— Mike