How we built ChatGPT’s novel ad stack for user trust and relevance

OpenAI built a bespoke advertising stack for ChatGPT to serve relevant ads without compromising user trust or the chatbot’s neutral voice. Unlike traditional request-response systems, ChatGPT’s conversational flow requires understanding evolving user intent across multiple turns and sessions. The system uses a post-trained large language model called the orchestrator, which makes tool calls for retrieval, ranking, relevance evaluation, and auction management. Key design choices include stateful ads sessions, a separate ads service running alongside ChatGPT’s response generation, and an independent auction mechanism. The orchestrator interprets conversation context, rewrites queries into semantic embeddings, retrieves a broad pool of candidates, filters them for eligibility, scores relevance, and nominates ads for auction. Final checks ensure contextual appropriateness and timely delivery. Ads are delivered server-side and displayed only if they arrive before the chat response completes. Logging and model-based evaluation create a self-improvement loop, while human feedback guides product direction. Within 200 days, this system achieved a $1 billion annualized revenue run rate with over 100,000 advertisers, using Codex to accelerate development.

(Source：OpenAI)