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

OpenAI published a detailed look at how ads are chosen inside ChatGPT. Instead of reacting to a single search, a custom post-trained model it calls the orchestrator tracks a conversation as it develops, summarizes the commercial intent, retrieves candidate ads, filters and scores them, and nominates finalists for an auction. OpenAI says ads run in a separate service with separate models and code from ChatGPT's answers, so the ad system can't shape the response.

The business scale is already large. OpenAI says ChatGPT ads hit a $1 billion annualized revenue run rate less than 200 days after launch, and more than 100,000 advertisers have run ads since.

Credit where due: the separation of ads from answers is the right call, and OpenAI explains it concretely. Different services, different models, no shared generation, and if the ad system is slow, no ad appears rather than the answer waiting. Returning no ad is treated as a normal outcome. That is a stronger design commitment than most ad platforms ever make, and it's the single most important thing for trust in an assistant people lean on for real decisions.

But the firewall protects the answer, not the user's attention or data. The ad session is stateful, remembering information across turns and, in OpenAI's words, across multiple conversations. When users opt into personalized ads, the orchestrator can also draw on what OpenAI calls nonsensitive information from past chats. Conversations are far richer than search queries; people work through health worries, money stress and family plans in them. Who decides what counts as nonsensitive, and how that line is audited, is the question this post doesn't fully answer.

There's also a subtler incentive problem. The answer can be untouched and the product can still drift. A business earning a billion-dollar run rate from conversational ads has every reason to favor conversations that produce commercial intent, and OpenAI says it plans more dynamic formats, including letting users talk directly to businesses through ads. The line between a helpful assistant and a sales channel is about to get tested.

One more thing stood out: OpenAI says Codex was a fundamental force multiplier in building the system so quickly, and that its strongest models evaluate the stack and train smaller ones in a self-improving loop. That is impressive engineering. It also means an advertising system that optimizes itself at machine speed, so the human feedback and published principles OpenAI cites have to carry real weight.

Our view: this is a thoughtful design for a business model with obvious pressure points. We'll judge it by how the privacy controls and the sensitive-topic exclusions hold up over time, not by the architecture diagram.

The flow runs from query rewriting and semantic retrieval through eligibility filters (budget, targeting, policy, pacing), relevance and engagement scoring, an auction that sets the price, and a final appropriateness check. Ads appear below the response, labeled as sponsored, with a short model-written preamble explaining how the ad relates to the conversation.