OpenAI Details the ChatGPT Ad Stack, With a Post-Trained LLM as Orchestrator

OpenAI's engineering team has published a detailed account of how it built advertising for ChatGPT from scratch. Conventional ad systems respond to discrete events such as a search query or a page view; in ChatGPT, intent unfolds across turns of a conversation, and some conversations never present a fitting advertising moment.

The result is a bespoke delivery system whose core is a custom, post-trained frontier model. It makes tool calls for retrieval, ranking and relevance scoring, then runs the auction that selects and prices the final ad.

Three design decisions define the stack. Ads sessions are stateful, so information persists across turns and even across conversations. The post-trained model interprets the conversation's context and underlying intent. And ad selection runs in its own service, fully separate from ChatGPT's answer generation.

That separation is more than architectural tidiness. OpenAI says the two services use different models and different code, so a slow or unavailable ads system cannot delay or take down the organic response. Returning no ad is treated as an ordinary, successful outcome, reserved for placements that clearly benefit both user and advertiser.

Timing decides whether an ad is ever seen. The frontend receives an ad creative bundled with a model-generated neutral preamble, and displays it below the answer, labeled as sponsored, only if it arrives before the response finishes generating. Keeping the preamble separate preserves the advertiser's message while holding ChatGPT's voice neutral.

The orchestrator does not pick the winner or set the price. It nominates contextually plausible candidates to the auction tool, which applies admission thresholds, budget pacing, bid and auction-feature computation, quality-adjusted ranking, pricing and final selection.

The commercial numbers are the hardest facts in the post: ChatGPT Ads reached $1 billion in annualized revenue run rate less than 200 days after launch, and more than 100,000 advertisers have run ads since launch. OpenAI also credits Codex as a force multiplier in building the system.

Improvement runs through models as well. OpenAI uses its strongest reasoning models to judge delivery outcomes and train smaller orchestrator and retrieval models, while clicks, hides, surveys and advertiser feedback through Ads Manager steer direction. People, the company says, shape taste more than the model's observational feedback does.

Next up are more dynamic formats, such as letting users talk directly to businesses through ads. The question that follows is whether users still trust the answer sitting above the ad.