How ChatGPT selects ads: OpenAI’s architecture explained

The useful bits

Ad selection and answer generation are separate processes.
Auction selection does not guarantee an impression.

How the ad-selection system works

Ritendra Datta and Karthik Gopalratnam’s 21 October engineering article explains ChatGPT’s advertising architecture. OpenAI describes a stateful ads session that follows evolving conversation context, with a separate service handling advertising while ChatGPT generates its response. An LLM orchestrator calls tools for candidate retrieval, eligibility, relevance, ranking and auctions. It proposes suitable candidates; it does not itself set the auction price. Winning selection is not the same as a displayed impression: context checks or a late response can still prevent delivery. Returning no ad is an intended outcome. The practical value of this guide is understanding those boundaries. It does not announce a new campaign switch, reveal a guaranteed ranking formula or promise that a particular creative tactic will secure placement. The accompanying diagram is OpenAI’s original illustration of the separate answer and advertising paths.

Questions for an advertiser’s review

Use the explanation to improve the questions asked during a test. Separate issues of account eligibility, creative relevance, available budget and observed delivery instead of treating every low-impression result as the same problem. That is an editorial diagnostic approach, not a claim that the platform exposes every internal decision in its reports. When explaining performance to stakeholders, distinguish the mechanics described by the platform from what the account data actually establishes. A technical account of how ads may be selected should not be presented as independent evidence of effectiveness. Keep the original engineering article as the reference for the architecture and the campaign’s own measurement plan as the reference for business outcomes.