OpenAI Reveals How ChatGPT Picks Ads — and Where Advertisers Stay Blind

ChatGPT Ads hit a $1 billion annualized revenue run rate less than 200 days after launch, with more than 100,000 advertisers running campaigns since debut. OpenAI shared those figures alongside its first technical breakdown of how the ChatGPT ad stack selects and serves ads. What the company did not share is any performance benchmark showing what advertisers earned from that spending.

How does the ad system stay separate from answers?

OpenAI says the system choosing ads runs separately from the model writing ChatGPT's answers, each with its own models and code. As ChatGPT composes a response, the ads system checks whether a relevant ad can appear below it. Choosing that ad does not feed back into what ChatGPT writes.

ChatGPT does not wait for the ads system. If ad selection is slow or fails, the answer still arrives, and a late ad never appears with that response. Each served ad carries a sponsored label plus a short system-written introduction, kept separate from the advertiser's headline, description, and call to action.

Separate models lend some support to OpenAI's promise that ads won't influence answers. This is OpenAI's own description of its architecture, not independent evidence the separation holds in practice.

Why does conversation context change ad matching?

The system reads beyond the latest question. OpenAI describes these as "stateful" sessions: each new message builds on details already shared, so the system never starts from scratch.

OpenAI's own example involves project management software. A user might explain they run a small agency, need client approvals, and want to avoid paying for guest accounts. A later pricing question then carries more context than a bare search for "project management software pricing." An advertiser offering free guest access could test that specific message against a generic productivity pitch — provided the landing page supports the offer.

With personalized ads enabled, the system can also draw on nonsensitive information from past conversations. That raises a harder problem: the model must distinguish current intent from stale information and hypothetical questions. Even a well-matched ad can reach someone whose purchase sits months away, so relevance alone says nothing about conversion likelihood.

What is the orchestrator?

OpenAI calls the coordinating model the "orchestrator." It converts relevant conversation parts into search queries, then into "embeddings" (numerical representations of meaning that match ads without identical wording). One conversation can produce several queries for different needs, and the orchestrator adjusts searches as information arrives.

Candidates then pass eligibility checks — location targeting, policy restrictions, budget, and pacing (the timing of budget spend). The system scores fit, past performance, and likelihood of a view, click, or conversion, then submits candidates to a separate auction tool that handles ranking, pricing, and selection. The language model does not choose the winner or set the price.

OpenAI did not disclose signal weights, so the post offers limited guidance on improving rank. Even auction winners face another relevance check against the specific conversation, and can still miss the delivery window entirely.

Why winning an auction doesn't guarantee an impression?

Low impressions can stem from several failure points: no matching ads found, failed eligibility checks, auction losses, or late delivery. Advertisers changing bids or creative without knowing where their ads stalled may waste the adjustment.

OpenAI says its internal logs record those outcomes. The post does not explain how much of this visibility advertisers get in their own reporting.

Does the revenue figure answer advertiser returns?

The $1 billion figure projects a full year from current pace; OpenAI has not collected $1 billion. The advertiser count omits how many stayed active, returned after testing, or increased spend — the closest proxy for whether advertisers found value. The post includes no CPA, ROAS, or conversion-rate benchmarks, no breakdown by industry or objective, and no controlled studies proving the ads generated incremental purchases rather than crediting buyers who would have converted anyway.

The takeaway: OpenAI's figures prove advertisers are spending; they do not prove advertisers are acquiring customers profitably. Advertisers testing ChatGPT Ads should diagnose low impressions before raising bids, and the industry should watch whether OpenAI discloses retention and cost metrics in the coming months.