On October 21 OpenAI did two things at once. It told Reuters that ChatGPT ads now run in 89 countries, 26 of them new, and it published an engineering post by Ritendra Datta and Karthik Gopalratnam on how the ad stack works. The expansion got the headlines. The post is the better document, because it is the first time anyone has written down, in public, how an advertising system behaves inside a product that is supposed to be acting for you.

The numbers first, since OpenAI put them in the post. ChatGPT Ads reached $1 billion in annualized revenue in under 200 days. More than 100,000 advertisers have run a campaign. Ads show only on the Free and Go plans. Plus, Pro and Enterprise stay clean. In September the company started testing business-sponsored agents with a handful of US advertisers, and the post ends by promising formats that let "users talk directly to businesses through ads."

I covered the launch in February as the moment the AI industry split on how it would get paid. This post is what the paid side built.

The ad system is a separate agent. It watches the conversation, remembers it across sessions, and bids on your intent. The answer you read is written by a different model that never sees the auction.

The wall between the answer and the ad

The design decision OpenAI leads with is separation. Ad selection "runs alongside ChatGPT's normal response generation," in its own service, with "separate models and separate code." The model writing your answer has no dependency on the ad system. If the ad service is slow or down, the answer arrives anyway. If the ad is not ready when the answer is, the ad is dropped for that turn. OpenAI calls returning no ad "an ordinary, successful outcome."

That wall is real and it is worth having. It is the difference between an ad next to an answer and an answer bent toward an ad. OpenAI's statement to the press is the plain version: ads are labeled, kept apart, and "advertising does not influence the answers ChatGPT provides."

Two details sit inside the wall, though. The ad system sends a "model-generated preamble," a short sentence explaining how the ad relates to what you were discussing, so the ad arrives with its own justification written by a model. And the ad system reads the same conversation the answer model reads. The wall stops ads from shaping answers. It does not stop the conversation from shaping ads. That is the point of the product.

The ad system has a memory the answer system does not use

Here is the part that should get more attention than the separation. A search ad responds to a query. OpenAI says ChatGPT "does not naturally produce such a concrete moment," so its ad system keeps "stateful ads sessions," which it defines as "information is remembered across turns of a conversation and across multiple conversations."

The worked example is a runner. The conversation starts with training questions. Several turns later it turns to shoes. The ad orchestrator, a custom post-trained model that calls retrieval, ranking and auction tools in a loop, is built to notice that the commercial meaning of the conversation changed, to tell a hypothetical from a purchase, and to try again with better queries as the conversation moves. If the user has opted into personalized ads, it also considers "available nonsensitive information about the user's past conversations."

In February Ben Thompson told John Collison that OpenAI's first ads were "terrible," and that the right model for an AI product was Meta's, not Google's: build a profile of the person rather than match the prompt, because profiling targets better and raises fewer questions about whether the answer was bent. Read the engineering post with that in mind. A session that remembers you across conversations is a profile under another name. OpenAI has built the Meta model and kept the Google-style wall. Thompson's objection was about the ads. His prescription is now the architecture.

Why this is a commerce story

The post describes an agent. It has a goal, it calls tools, it iterates, it is judged by larger models and retrained on the results. Its goal is to find "the ad that balances high user value and high advertiser value." OpenAI says the orchestrator puts users first. I take the sentence at face value and still note that the balance is set by OpenAI and the auction, not by the person asking about shoes.

That matters because the same company is running sponsored agents and promising ads you can talk to. I wrote in September that the shortlist is for sale: discovery inside the conversation now has a price and a limited number of slots. This post is the machine that fills the slots. The next step the authors name, talking to a business through its ad, is the point where an ad stops being a placement and becomes a counterparty. Then the question is not whether the ad influenced the answer. It is whether the thing you are now negotiating with was chosen by an auction.

In the MM Destination Economy Thesis, the shelf moves inside the conversation and the merchant's own storefront loses value. OpenAI's post is the first public account of how that shelf is stocked: by a model with a memory, a bid, and a sentence of its own to explain why it is there.

What a merchant and a buyer should take from it

For an advertiser, the useful fact is that intent is modeled over the whole conversation, not the last message. A campaign that only wins on keyword matches will lose to one that fits the goal the system has inferred. AdExchanger reports advertisers are asking for measurement before they commit more budget, and OpenAI says it is building it alongside the expansion. Until then the attribution loop is the advertiser's own.

For anyone shipping an agent that shops, the post is a design reference with the roles reversed. OpenAI built a separate, auditable agent for the party that pays, and gave it memory and a bid. The buyer's agent, the one that is meant to act for the person, mostly does not exist yet at this level of engineering. When it does, it will meet this one at the shelf.

The paid conversation is live in 89 countries. The write-up of what it remembers about you took one engineering post. The write-up of what a buyer's agent is allowed to remember, and who it works for, has not been written by anyone.

If the ad system remembers your last ten conversations and the answer system is kept from using them, which of the two is the one acting for you?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.