ChatGPT picks ads with a separate model from the one that writes answers, OpenAI says

OpenAI has published how ChatGPT picks and places ads, and the central design choice is a wall. Ad selection runs in its own service, with separate models and separate code, alongside the answer ChatGPT generates. OpenAI says the boundary exists "to keep ads from influencing answers and maintain ChatGPT's neutral voice."

The OpenAI engineering post, by Ritendra Datta and Karthik Gopalratnam, also gives the scale. ChatGPT Ads reached $1 billion in annualized revenue run rate less than 200 days after launch, and more than 100,000 advertisers have run ads on ChatGPT since launch.

How the ad stack works

• A custom, post-trained LLM, which OpenAI calls the orchestrator, runs the ads flow through tool calls for retrieval, ranking, relevance and auctions.

• Ads sessions are stateful. The system remembers context across turns and across conversations, so it reads intent from the whole exchange.

• The orchestrator condenses the conversation into one or more short queries, turns them into embeddings and retrieves candidates from several indices.

• Candidates are filtered on budget, pacing, targeting, policy and placement compatibility before final ranking.

• Ads are placed at the end of a chat response.

• In OpenAI's words, returning no ad is "an ordinary, successful outcome."

Two surfaces, two jobs

If the separation works as described, a brand in ChatGPT is now measured on two independent things: whether the model names it in the answer, and whether an ad appears under that answer. Budget buys the second. By OpenAI's account, it has no route into the first.

That cuts both ways. A brand the model never recommends cannot pay its way into the answer. A brand the model already recommends will likely share the screen with more paid competitors over time, as more advertisers join a channel already past $1 billion of run rate.

Reviews feed the paid side too

One detail matters for anyone selling products. OpenAI says the index for product feed ads "is designed to be aware of positive or negative product reviews" when it fetches ads. So reputation is an input to paid retrieval as well. A weak review profile can cost a brand its ad slot as well as its mentions.

What the post does not say

The post does not explain how auction prices are set or what advertisers can see about the conversations their ads appear in. It also gives no independent check that the boundary holds in practice. It is a description of the system by the team that built it.

What to do

Track the two surfaces separately. Organic mentions tell you whether the model trusts you. Ad presence tells you who is buying the conversation around you. A ChatGPT plan that covers only one of them leaves half the screen unwatched.