On May 5, OpenAI opened the ChatGPT Ads Manager to every US advertiser with no minimum spend. That sentence is shorter than the change it describes. For most of the last three years, the question “what does it cost to put an ad inside the answer box” did not have a numeric answer. As of this month it does. The published CPM range is twenty-five to sixty dollars. The CPC floor is three to five. The platform reportedly cleared about a hundred million dollars in its first six weeks and is targeting two and a half billion this year. Those are not press release numbers anymore. They are media buying numbers. The implications get cleaner once they are treated that way.
The February rollout was the soft launch. Sponsored answers showed up under the AI response, clearly labeled, available only on the US Free and ChatGPT Go tiers, and the inventory was managed by direct-sold deals with a small set of brands and agencies. The May change is the difference between a private auction and a public one. Self-serve means a sole proprietor in Dayton can buy the same inventory slot underneath a ChatGPT answer that Dentsu, Omnicom, Publicis and WPP are now integrating into their managed-service offerings. The number that should have been newsworthy in the May coverage was not the launch itself. It was “no minimum spend.”
The CPM math, plainly
A twenty-five to sixty dollar CPM is not, on its own, a useful number. The question is always: against what response rate, and compared to what alternative. Search ads on Google clear in the low double digits for most categories. Display CPMs on the open web settle in the single digits. Meta’s auction-cleared CPMs sit in a similar range to ChatGPT’s lower bound, sometimes higher in saturated verticals. So the headline pricing is not, by itself, evidence of either a bubble or a bargain. What changes the math is the contextual fit.
The ad slot underneath the AI answer is shown to a person whose intent is visible in a way that a search query does not capture. A search query is a phrase. The conversation that produces the prompt is a sequence: the user said why they were looking, what they had already considered, and what they did not understand yet. The CPM is higher than display because the placement is closer to a decision than display can get. It is competitive with search because the intent signal is comparable. Whether that intent signal converts at the rate the pricing implies is the question the next two quarters of campaign data will answer.
We have looked at the early case studies floating around the marketing press and they are not yet trustworthy. The advertisers who participated in the February-to-May invite-only phase had hand-tuned creative, account-management support, and the novelty advantage of being among the first dozen brands a user had seen show up there. None of that holds for the May-onward cohort. The first set of self-serve campaigns will be the honest read.
Targeting that is not keyword targeting
The targeting model is the part of the announcement that has not been absorbed yet. ChatGPT does not match ads against keywords. It matches against the inferred topic of the conversation, the user’s prior chat history where the user has not opted out, and prior ad interactions inside the platform. That is closer to a recommender system than to a search auction. The advertiser does not buy “the phrase running shoes.” They buy “the conversational context our model thinks is a running-shoes intent.”
This has two consequences that we think will land harder than the launch coverage suggested. The first is that the bidding strategy that works on Google or Meta will not directly transfer. The unit of targeting is not a keyword set or an interest cluster. It is an intent inference produced by a model whose internals the advertiser cannot inspect. The campaigns that learn the placement fastest will be the ones that treat the model’s intent labels as a black box to be tested against, not as a taxonomy to be planned against.
The second is that contextual targeting against a conversation is materially more invasive than contextual targeting against a search query, even before the chat history layer is added. A search query is a hundred and twenty characters at most. A chat is a transcript. The amount of detail the targeting layer can in principle act on is much larger. Whether the platform exposes that detail to advertisers as targeting controls, or hides it behind opaque “interested in fitness gear” buckets, is a policy choice OpenAI has not finished making in public. The May announcement promised third-party measurement and CPA bidding. It did not promise transparency on what the intent layer is actually inferring.
The trust question, sharper
The argument we have seen most often this month is the one about whether ads will bias the answer. We do not think the bias risk is the most interesting version of the question. The format keeps the sponsored block visually separate from the AI response and labels it as advertising. The model is not, on any reading of the OpenAI policy documents, supposed to alter its answer based on which advertiser bid on the conversation.
The more durable trust question is about the inversion of the relationship. For three years, the value proposition of ChatGPT was that the model worked for the user. The user paid in subscription dollars or in the cost of the free tier’s compute, and the model gave the most useful answer it could compute. With an ad inventory layer, the model still works for the user, but the platform now also works for the advertiser, and the advertiser is paying more per impression than the free user is. The incentives do not collide on any single answer. They collide on the slow product decisions: which queries get richer response surfaces, which categories get more granular intent labels, which experiments get prioritised. Those decisions are not visible in a single conversation. They are visible only across the next eighteen months of product changes.
What the next eighteen months tell us
The number to watch is not the two and a half billion revenue target. It is the steady-state ratio between subscription revenue and ad revenue, and which side of that ratio OpenAI uses to justify each product decision. The platform is currently making a public case that ads exist to fund free access. That case is coherent on the May numbers. It will stay coherent as long as the subscription business stays the larger half. The first quarter where ad revenue exceeds subscription revenue is the quarter where the same case stops being coherent without a different framing.
The May 5 launch is the moment the question stopped being theoretical. Everything since has been the first round of pricing discovery, and the pricing has been telegraphed. The second round, the one that decides whether the format settles in or gets recoiled from, is the one we will be reading about in the fall.