OpenAI Details ChatGPT Ad System That Keeps Ads Apart From Replies

OpenAI’s ChatGPT ad system uses a post-trained language model to interpret chats and nominate ads; separate tools score candidates and an auction chooses placements and prices, apart from answer generation.

The system converts conversation context into queries and embeddings, or numerical representations, then uses approximate-neighbor indexes to find possible ads. Separate tools assess each candidate’s eligibility, relevance and engagement. Filters apply targeting, policy, budget and pacing rules, and remove duplicate ads.

The language model’s role is to nominate candidates. An auction tool makes the final selection and pricing decisions. The ad stack runs on separate services, models and code from ChatGPT’s answer generation, so a slow or unavailable ad service does not hold up a response.

A final check can suppress ads judged inappropriate or too late to display. Sponsored placements appear below responses with advertiser creative and a neutral preamble generated by a separate model. Campaign measurement and optimization use Pixel and Conversions API tools.

The technical details add context to TokenPost’s coverage of ChatGPT ads expanding to 89 countries. The advertising business reached a $1 billion annualized revenue run rate on Aug. 31, a pace measure rather than $1 billion in realized annual revenue.