OpenAI has published a technical account of how ads reach users inside ChatGPT, describing a dedicated language model that reads a conversation, finds candidate ads, and passes them to a separate auction. Search Engine Journal, in a report by Brooke Osmundson on October 22, summarized the post and flagged what it does not say.
For search teams, the stakes are practical. Paid placements now sit next to AI-written answers, and the mechanics decide how advertisers diagnose weak delivery and whether users can trust the answer itself.
OpenAI says the system that chooses ads runs apart from the system that writes responses, with its own models and code. The answer does not wait on the ad. If the ad system is slow or fails, the response still arrives, and a late ad simply misses that turn. Ads carry a sponsored label under the response, plus a short introduction tying the ad to the discussion.
That separation is OpenAI’s description of its own design. The post offers no independent test showing that ads never shape answers. Treat it as a stated architecture, not a verified guarantee.
The matching logic reaches back through the whole chat. OpenAI calls these sessions “stateful,” so earlier details carry forward instead of resetting with each message. Someone who mentions running a small agency and avoiding guest-account fees gives the system more to work with than a bare pricing query would. Users who opt into personalization add another input: nonsensitive details from earlier chats. That forces the model to tell live needs apart from stale or hypothetical ones.
The model that manages selection is called the “orchestrator.” It rewrites pieces of the chat as lookup queries, encodes those as embeddings (numerical representations of meaning), and looks up ads without needing matching wording. One conversation can produce several queries.
Candidates then face eligibility checks, including location targeting, policy limits, budget, and pacing. The orchestrator scores each ad for fit with the chat, weighs its track record, and estimates the odds of a view, click, or conversion. It sends the shortlist to an auction tool that sets ranking and price. The language model decides who enters, not who wins.
OpenAI did not disclose how heavily each signal counts. That makes the breakdown a map of the stages, not a guide to ranking higher.
Winning is not the end. A further check judges whether the selected ad suits that particular conversation, and an ad can pass policy review and still be held back. An auction winner may also arrive too late to be shown. Low impressions can therefore come from relevance filters or timing, not only from bids or creative.
OpenAI says internal logs record where ads drop out: searches that found nothing, ads that failed eligibility, and placements that drew an interaction. The post does not say how much of that reaches advertisers. Without it, a campaign blocked by location targeting looks identical to one that enters auctions and rarely wins.
The business figures come with gaps. OpenAI says ChatGPT Ads hit a $1 billion annualized run rate within 200 days of launch, and that more than 100,000 advertisers have run ads. A run rate projects a year from current pace; it is not money already collected. Neither number reveals retention, spend growth, cost per acquisition, return on ad spend, or whether the ads produced sales that would have happened anyway.
Compare that with Google Ads, where advertisers are used to diagnosing delivery through auction and impression-share reporting. ChatGPT Ads has described its internal funnel without showing buyers an equivalent view, at least in this post.
Advertisers already running ChatGPT Ads should ask OpenAI for stage-level reporting before raising bids to fix low impressions. Teams still deciding on budget should hold test spend modest until the company publishes conversion data that someone other than OpenAI can check.
Based on reporting by Brooke Osmundson for Search Engine Journal, published October 22, 2026, which covered OpenAI’s technical post on its ChatGPT ad system.
