The conversation is the inventory now

OpenAI published a short blog post on May 7 confirming what had been trailed for months: ChatGPT now carries ads. The Ads Manager beta is live at ads.openai.com. The previous $50,000 minimum spend is gone. Cost-per-click bidding is in. Lifestyle, travel, education, and digital-products advertisers can buy placements that appear, labelled "Sponsored," beneath responses for free-tier users. Paid tiers stay ad-free. OpenAI says it crossed $100 million in annualised ad revenue within six weeks of the earlier closed pilot.¹

The framing in the press coverage has been mostly about business model, OpenAI diversifying revenue, the inevitability of ads in any free consumer product, the comparison to Google circa 2001. That framing is not wrong, exactly. It's just narrow. What's actually happening is that one of the most intimate interfaces a person has ever had with software, a system you talk to in natural language, that holds your context across sessions, that you increasingly use to think, is being wired into an attention market.

I want to take that seriously, because I don't think we have, collectively, sat with what it means.

What the search-ads parallel misses. Google's ads, then and now, sit beside a list of links you scan and choose between. The user is doing the synthesising. The ad is one item in a visible set, and you can see, roughly, what's an ad and what isn't. ChatGPT does not work like that. ChatGPT gives you an answer. It composes prose. It picks, on your behalf, what to mention and what to leave out. When a sponsored item enters that flow, even labelled, it enters as part of a synthesised recommendation from a system the user has been trained, by the product itself, to treat as an assistant rather than a salesperson.

This is not a small distinction. The Federal Trade Commission has been clear since at least 2013 that the line between editorial and advertising content has to be unmistakable.² "Sponsored" labels on a list of links work because the user is already in evaluation mode. "Sponsored" labels on a recommendation embedded inside an answer ask the user to do cognitive work the interface is otherwise designed to remove. The whole pitch of ChatGPT is that you don't have to evaluate ten links, it does it for you. Now some of what it does for you is paid placement.

Who pays. The paid tier is ad-free. That sentence is doing a lot of work. It means the people who can afford $20 or $200 a month get the unmediated assistant. The people who can't get the assistant whose recommendations are now partly shaped by who paid to be in them. We have a name for this arrangement when it shows up elsewhere: a two-tier information environment, where the quality of what you're told depends on what you can pay.

That number matters because it tells you how fast this scales. Six weeks of a closed pilot, with a $50,000 minimum and a narrow advertiser pool, produced a nine-figure annualised run rate. The Ads Manager beta has now removed the minimum and opened the categories. Whatever you think the equilibrium looks like, it's bigger than the pilot, and it arrives quickly. The economic gravity here is not subtle. Once a product can monetise at this rate against an audience already in the hundreds of millions, the question of whether ads expand is settled. The only open questions are how much, where, and how visibly.

The categories are not random. Lifestyle, travel, education, digital products. Read that list again. These are the categories where consumer trust in a recommendation translates most directly into purchase, and where the marginal user is most often making a decision they don't have deep expertise in. Someone asking ChatGPT to help plan a trip, pick a course, or choose a productivity tool is exactly the user for whom a synthesised recommendation has the most weight, because they came in without strong priors. Education is the one that should give us the longest pause. A free-tier student asking ChatGPT which online course to take, which certification matters in their field, which bootcamp is worth the money, is now asking a system whose answer is shaped, in part, by who paid for placement in that vertical. The information asymmetry between the student and the advertiser was already large. The interface just narrowed the student's view of it.

The consent question. Users were not asked. There was no opt-in moment, no period of public consultation, no published methodology for how sponsored content interacts with organic recommendations, no disclosure of what advertisers can and cannot influence. OpenAI's blog post is a product announcement, not a policy document. We do not know, as users or as a public, whether an advertiser in the travel category can bid to appear in answers to general "where should I go for a week in October" queries, or only when the user names a brand. We do not know how the ranking interacts with what the model would otherwise say. We do not know whether the conversational context, the things you've told the assistant about yourself across sessions, informs ad targeting. The blog post does not say. The Ads Manager documentation, from what's public, does not say either.

This is the thing I keep coming back to. The interface collects more about you than any previous consumer ad surface has ever collected. Not browsing behaviour. Not search queries. Conversation. The things you're worried about. The decisions you're weighing. The personal context you've fed in to get better answers. That data, in an ads regime, becomes targetable, unless the company commits, publicly and verifiably, that it doesn't. OpenAI has not made that commitment in the May 7 post. The absence is doing work.

The interface collects more about you than any previous consumer ad surface has ever collected. Not browsing behaviour. Conversation.

The historical pattern. Every consumer information platform that has gone from subscription or free-with-no-ads to ad-supported has followed roughly the same trajectory. Ads start labelled and contained. The labelling erodes. The placements expand. The targeting deepens. The editorial product bends, slowly, toward what monetises. This is not a conspiracy theory; it is the documented history of free-to-air television, commercial radio, web search, social feeds, and podcasts. The pressure is structural. The advertiser is the customer; the user is, in the old phrase, the product. Whether OpenAI bends in the same way depends on choices it has not yet committed to publicly, and on regulatory pressure that does not yet exist for this product surface.

I want to be careful here. I am not arguing that ads in ChatGPT are uniquely evil, or that OpenAI is uniquely cynical. The economics of running a model of this scale are real. Free-tier users do cost money to serve. Subscription alone, at the prices the market will bear, may not cover it. The question is not whether OpenAI needed revenue. The question is what was traded for it, and who was at the table when the trade was made.

What's actually being traded. Three things, mostly. First, the trust premium that made ChatGPT useful in the first place, the sense that the answer you get is the system's best attempt at the answer, full stop. Second, the cleanness of the conversational interface as a space relatively uncontaminated by commercial intermediation, which made it valuable for thinking, drafting, working through problems. Third, the integrity of the recommendation function for the people who can't pay to opt out, which is most users. None of these things have been destroyed by the May 7 announcement. All three have been put up for incremental erosion, on a curve set by ad revenue growth.

What to watch. A few things, concretely. Whether OpenAI publishes, soon, a clear policy on what advertisers can and cannot influence, not a marketing page, a policy document with specifics. Whether the labelling stays unambiguous as placements expand into more categories. Whether conversational context becomes a targeting input, with or without disclosure. Whether the FTC or its European counterparts issue guidance specific to AI-assistant advertising, or wait, as they did with social platforms, until the harms are documented and entrenched. Whether the paid tier remains genuinely ad-free, or whether "ad-free" becomes "fewer ads" the way it has on every other tiered consumer service.

And one thing more. Whether researchers get access to study any of this. The closest analogue we have for understanding how recommendation-with-ads shapes user behaviour at scale is social media, and the answer to "did independent researchers get good access" was no, for over a decade, until the harms were already in the system. If we run that experiment again, on a more intimate interface, with less independent oversight, we will learn the same lessons later and at higher cost.

The harder thing. It would be easier to write this piece as either "ads in ChatGPT are fine, the labelling works, calm down" or "ads in ChatGPT are the end of trustworthy AI, this is a catastrophe." Neither is true. What's true is that a structural change has been made, quietly, to the most-used AI interface on the planet, on the basis of a product blog post and an ad sales page. The change is small now. The economic logic guarantees it will not stay small. The people most affected, free-tier users, who skew toward those who can't or won't pay, which globally means most of the world, were not consulted, are not represented in the design, and will absorb most of the cost in attention they don't notice they're spending.

I find it hard to write about this without sounding like I am performing concern, so let me be direct. I think this is a worse outcome than the one we had on May 6. Not catastrophically worse. Not irreversibly worse. Just worse, in the specific sense that a tool I and many others have used to think with is now also, partly, a tool used to sell to us, and the people who can pay to keep it as the first thing have been quietly separated from the people who can't.

That separation is the story. The rest is implementation detail.

Footnotes

¶