ChatGPT Opened 65 Pages to Buy a Shower Head

Sonya Huang introduces James Cadwallader as co-founder and CEO of Profound, a marketing platform for the AI era. The company helps brands see how they show up in ChatGPT and Claude, what to do about those rankings, and how to give a single marketer the power of an agency. She says they serve 10% of the Fortune 500. Cadwallader corrects her on the air: it is close to 12% now, across finance, consumer, CPG, software, and both B2B and direct-to-consumer. ChatGPT is rolling out ads, and every marketer is trying to live in what Huang calls the brave new world of agent-led growth.

Cadwallader calls this the biggest platform shift in the history of marketing: from predetermined blue-link search to probabilistic AI answers. Showing up when someone asks ChatGPT about your category is only the start. What does ChatGPT say about you? How does Claude recommend your coding tools? Agents and superintelligence, he argues, are replacing the role of the consumer. The front door of the internet is not what changed. The person going through the door is.

It used to be a shopper with a list of blue links, clicking into your site. Now an agent uses a similar index, discovers your brand, products, and services, then comes back through the door and keeps the relationship with the human. Huang restates the point: ChatGPT may be the new front door, but the important fact for marketers is the agent walking through it and making a large part of the buying decision. Ask ChatGPT or Gemini a question, Cadwallader says, and you are asking an agent to crawl the internet on your behalf, visit the sites you used to visit, decide what is useful, and remix it into an answer.

The biggest misconception, Cadwallader says, is that this is just SEO. The misconception exists for a reason. Brands still have to create content and distribute it on owned, earned, and social channels, and ranking on an index still matters. The human consumer, though, is no longer the one using the web. You may be building content that is never consumed by a person at all. In SEO, content was designed to be picked up by an algorithm and then read by a human. In this world, it is designed to be both discovered and consumed by an agent.

Humans are impatient, and that used to be the whole game. In classic SEO, 95% of the value sat in the top four or five blue links, a function of scarce cognitive energy, scarce patience, and scarce time. Agents use the index differently. ChatGPT, Gemini, and Claude are far more willing to work the long tail, and the surface area they will use to answer a question is orders of magnitude wider than a human's. Three or four months before the taping, Cadwallader asked ChatGPT to help him find a shower head for his New York apartment. It used 65 different web pages.

Huang jokes that she probably would have opened 65 pages herself for a purchase that important. Cadwallader's point is the marketer's: you are building for a superintelligent agent with infinite bandwidth. As the cost of inference falls, those agents will only use more of the internet to assemble rich answers. High-consideration, high-ticket categories feel it first, consumer electronics, cars, white goods, anything that already demanded research, because AI is unusually good at that work. Huang, as it happens, is using AI to buy a car while they talk.

ChatGPT, Claude, Gemini, and Grok recommend brands in wildly different ways. Cadwallader's shorthand is to treat them as different species, which is also why software like Profound exists. The product looks at how often you appear on each platform, the sentiment and neighboring themes when you do, and then the citations each model uses to answer questions about your brand, category, or competitors. Once you have the what and the why, the next step inside Profound is to deploy your own agents to sort that data and build a new kind of marketing.

The roots of the differences are practical. Gemini leans heavily on YouTube, which makes sense because Google owns YouTube, so video is a huge lever if you want to appear in Gemini answers. ChatGPT typically pulls from Reddit on consumer questions and from LinkedIn on B2B, which Cadwallader calls a huge source of truth. Claude used to lean more on the pretrained model and is now, after classifier updates, more sensitive to real-time information and more willing to use the web. Huang has watched recommendations shift even from Claude 4.5 to 4.6. Humans, Cadwallader says, are creatures of heuristic; they want to pattern-match this to old search. Huang's phrase is ALG, agent-led growth. Claude is not merely a new discovery channel. Claude is a user. If you are vibe-coding with Claude Code, does it recommend MongoDB or Vercel, why, and where did it learn that?

Agents, he has told Huang, will consume a hundred times more internet, which sounds like a mandate to spam a landing page for every long-tail shower-head query. He thinks "AI slop" is a red herring. A recent estimate put about half the web on AI-written content, and a *New York Times* blind test found 53% of readers preferred the AI article to the journalist's. AI can write high-quality marketing. The real task is that the consumer is now superintelligent, so you have to tell Claude something it does not know. Humans, he says, are a fleshy API between reality and the internet. Huang answers on cue: she is just a fleshy API. For a new Nike Alphafly, the job is first-principles marketing, equipping the model to answer any question about the product without poisoning it. Comparative listicles still work, he admits, because models prefer content that already did the hard comparison, Sequoia ranked first, weaker names below, real rivals omitted. He would not advise it, and he already prompts models to ignore listicles and brand-published pages. As first-principles reasoning gets stronger, that shortcut should get punished.

Huang asks about dead-internet theory: when agents do most of the browsing, does a marketing site still matter, or should everyone just publish a README for crawlers? Cadwallader is willing to sound dramatic. He thinks a dead-internet outcome in the next three years is possible, not likely, but possible. If humans only talk to AI to get answers, the incentive to publish falls toward zero. Both people and models still depend on first-party reporting, and most of the internet is still funded by advertising. When humans stop clicking into pages, the ad model breaks, editorial economics break, and then you have to ask where AI goes to find answers.

A brand README can work for commercially driven content. Huang imagines a page that says she is Sonya, here are the companies she has backed, so a founder's agent can find her. Cadwallader agrees for that use case. Much of the internet, though, is people yapping and sharing ideas, and that machine ran on money and status. Publish a good blog, run ads, get paid, or get famous. If AI vacuums out the good parts, remixes them, and leaves a citation nobody clicks, the fleshy API stops bothering. His second-order bet is that every AI lab eventually vertically integrates with a social network. Social media becomes more human: Meta already makes it hard to bot-post on Instagram; X and YouTube are starting to pay for good work. Grok already uses X's data with some skill. Reddit is embracing its humanity as a scarce asset and has a large deal with OpenAI. Purely on vibes, he has wondered whether OpenAI might acquire Reddit.

Could platforms simply ban scrapers and sell access, a revenue share back to original creators? X has opened its API, Reddit already has the OpenAI deal, and Cadwallader reads both as versions of that vertical integration. The other way off the fleshy-API path is robots. If the world ends up with 50 billion of them, bipedal or drones, AI can capture first-party data and understand reality without waiting for a human to type it up.

Ads are arriving inside generative agents, and marketers worry users will revolt. Cadwallader thinks people will get over it quickly, as they did with Google. Generative advertising in a conversational interface, with high personalization, will be the most effective form of advertising the world has ever seen. The conversations already hold rich consumer intent. The model also understands Sonya, in that exact moment, well enough to append a super-personalized ad at the point she would be most responsive. He has a team working on ads.

The unit itself, he says, is not his original idea; OpenAI has alluded to it. An advertiser will system-prompt a campaign. Target women in Minnesota between 35 and 40, mention these points when they talk about photography, do not mention those, use this brand knowledge base, match their tone of voice, and aim for this much revenue. B2B and coding agents are a different problem, because the agent is steering the purchase and people want that agent objective. If Claude picked a database because of a good ad, you would tell it to use whatever is best.

Humans, Cadwallader notes, are not objective creatures. We just like to think we are. You have already been swayed by Vercel's branding and by a podcast with MongoDB's founder, and you do not know it. That is why we demand objectivity from agents. The relationship between advertising and agents shopping on our behalf, he says, is going to be quite different from the old banner-and-search world.

Huang asks for parting wisdom for marketers whose bosses send screenshots asking why they are not number one. Cadwallader's first answer is to use Profound. He is only half joking. Without a platform that shows how you actually appear, you are guessing. Reductionisms like "LinkedIn matters" or "Reddit matters" will fail if you run on vibes, because each category has its own sources. You have to look at the citations to see how and why a model mentions you.

The second part is harsher. Marketing teams that are not using agents, and in particular Profound's agents, to build and distribute marketing are failing. It has gone from a nice-to-have to a must-have. The misconception is that agents merely automate yesterday's work. Cadwallader says the reality is a type of marketing that was not possible before. Profound built an internal agent that reads Gong transcripts from sales calls, captures objections, buckets them into themes, and writes battle cards from a product knowledge base for the sales team in real time. Eighteen months earlier that would have been a human doing the job once a quarter. Now it runs every day.

Huang restates the advice: get visibility, because you cannot optimize what you cannot see, and fully embrace agentic marketing or you are behind. Cadwallader's analogy is refusing the internet. It is like saying you will stick to print and television, thanks, you do not really believe in this internet thing.