ChatGPT ads hit $1 billion run rate in under 200 days

OpenAI frames the core problem as one of timing rather than targeting. Traditional ad systems respond to a discrete event such as a search query or page view, but a ChatGPT conversation may carry commercial intent that only becomes clear after several turns, as in its running example of a user who asks about race training and only later asks about shoes. Building around single messages, OpenAI says, risked misreading the goal, starting the ad flow too soon, or rerunning the flow on every change.

Its answer is a stateful ads session plus an orchestrator LLM that summarizes conversations, retrieves candidates, filters for eligibility and relevance, and nominates candidates for an auction without itself setting the price. OpenAI draws a hard architectural line between this flow and answer generation, and it notes the practical upside: ad serving issues cannot take down or delay ChatGPT, and delivering no ad counts as a successful outcome.

The build itself leaned on Codex, which OpenAI describes as a force multiplier for bootstrapping the system and its guardrails at speed. OpenAI credits human feedback, including clicks, hides, surveys and advertiser input through Ads Manager, with shaping taste more than the model's own observations. The next stated step is more dynamic formats, including letting users talk directly to businesses through ads.