OpenAI explains how ChatGPT picks and serves ads

OpenAI has published a technical breakdown of how ads reach people inside ChatGPT. Search Engine Journal summarised it on 22 October 2026. We could not open OpenAI's page ourselves, so the details below follow that report.

The main point: ad selection runs separately from answer generation, with its own models and code. The ads system looks for a relevant ad while the answer is written, but the ad choice does not change the answer. If the ads system is slow or fails, ChatGPT still shows the answer, and a late ad simply does not appear.

How an ad gets chosen

According to the report, the process has several steps:

1. A model called the orchestrator reads the conversation and calls tools to find candidate ads.
2. It turns relevant parts of the conversation into search queries, stored as embeddings so ads can match on meaning rather than exact words. One conversation can produce several queries.
3. Candidates are checked for eligibility: location targeting, policy limits, remaining budget and pacing.
4. Eligible ads are scored for fit with the conversation, past performance and the likelihood of a view, click or conversion.
5. A separate auction tool handles ranking, pricing and the final pick. The language model helps choose entrants but does not pick the winner or set the price.
6. A last check asks whether the winning ad fits the conversation, even if it meets ad policies.

Ads appear below the response with a sponsored label. The system writes a short introduction that links the ad to the discussion, kept apart from your headline, description and call to action. OpenAI did not say how much weight each signal gets.

Why it matters

The system tracks details across messages, not only the latest question. The report gives the example of someone comparing project management tools who mentions running a small agency and wanting to avoid guest-account fees. When users choose personalised ads, the system can also use nonsensitive information from past conversations.

For advertisers, that means targeting follows the full conversation. Messaging that speaks to a specific need, such as a constraint or a use case, may match better than broad claims.

It also means a low impression count can have several causes: no candidate found, a failed eligibility check, a rejected fit check or a missed delivery window. OpenAI says its internal logs record where ads stop, but the report does not say how much of that advertisers can see.

The report also cites two figures from OpenAI: a run rate of $1 billion a year in under 200 days since launch, and more than 100,000 advertisers since launch. Those show spending, not results. The post, as summarised, includes no cost per acquisition, return on ad spend or incrementality data.

What to do this week

• List the real constraints and use cases your buyers state before they choose you. Write ad copy and landing pages around them.
• If you test ChatGPT Ads, judge results on lead quality and revenue, not clicks. Run a manual campaign next to any automated one.
• Ask your account contact what reporting exists for ads that were eligible but not shown.
• Remember that OpenAI plans to turn on creative enhancements by default on 30 October. Read our report on that change and decide per campaign.
• Check which of your organic pages ChatGPT already cites, since paid and organic often reward the same pages.

Aiso shows the questions buyers ask AI tools and whether your brand appears in the answers, which helps you write for real conversation context. If ChatGPT clicks land on your site but few turn into leads, Javo can answer visitors' questions on the page and capture their details.