- OpenAI says its ChatGPT ad system uses a trained AI model to read conversations over time, then runs auctions to pick ads.
- OpenAI says ads run in a separate service to keep ads from shaping answers. That is the company's description, not an outside finding.
- Advertisers and security teams should ask how intent is inferred, what 'delivered' means, and how personalization applies to their staff.
From a question to a direction
For years, online ads followed a simple trigger. A person typed a search or opened a page, and the ad system matched that event to an ad. OpenAI's engineers say ChatGPT does not fit that model. In a post published October 21, Ritendra Datta and Karthik Gopalratnam describe an ad stack built from scratch.
Their reason is that people arrive with goals, not one-line queries. They plan trips, pick software or learn a hobby, and the need sharpens over many turns. The authors give an example. A chat starts with a general question about running. Several turns later, the user asks about shoes. Acting on the first message could, in their words, retrieve "the wrong intent entirely."
This shows a shift in what is being sold. Search ads sell a query. ChatGPT ads, as OpenAI describes them, sell a reading of where a person is heading. That moves the trust question from "what did you ask?" to "what does the system think you want?"
How the machine picks an ad
OpenAI calls the central piece an orchestrator. It is a custom, post-trained language model that calls other tools in a loop. It does not choose the ad alone.
First, the orchestrator turns the conversation into short text summaries. Those become embeddings, which are stored meanings a system can search quickly. The post gives a hiking case. A user asks what to cover on a two-day Yosemite trip and later asks about a shoe brand. The system might form two queries: one for shoes suited to a hiking trip, and one for the brand's products good for travel.
Next comes retrieval. OpenAI says it builds several indices and uses approximate-nearest-neighbor search. That method trades perfect answers for speed, while keeping most relevant candidates.
Filters then remove ads that fail checks. These include campaign targeting, policy rules, remaining budget, pacing and limits on how many ads one advertiser gets. Scoring tools rate semantic relevance, predicted engagement and past performance.
The orchestrator then nominates plausible candidates. OpenAI states that it does not pick the winner or set the price. A separate auction tool does that.
The wall between ads and answers
OpenAI's central trust claim is architectural. The ads system runs in its own service, with separate models and separate code. The authors say the aim is to keep ad logic from shaping what ChatGPT says, and to preserve its neutral voice.
It also has an operational effect. If ad serving is slow or down, OpenAI says the organic response continues. The post adds that returning no ad is "an ordinary, successful outcome."
This is OpenAI's own description. The post offers no outside audit or test of the boundary. Leaders should treat it as a claim to verify, not a finding.
What 'delivered' means
The post is candid about gaps between delivery and display. After the auction, the orchestrator can block an ad that feels wrong for the conversation. An ad can also miss the response and time out, because OpenAI says it does not delay chat answers to wait for ads.
So server-side delivery does not guarantee that a user sees an ad. For anyone buying, that distinction belongs in reporting.
OpenAI also gives two business figures. It says its ad product reached an annualized revenue pace of $1 billion within 200 days of starting. It adds that over 100,000 advertisers have used ChatGPT to run ads. Both numbers come from the company, and the post gives no breakdown.
Questions to put to your teams
For marketing: ask how conversations map to the intent you are paying for. Ask whether reports separate ads delivered from ads displayed.
For security and privacy: the post says that if a user chooses personalization, the orchestrator considers nonsensitive information from past conversations. It does not say how this applies to business accounts. Ask what your staff's settings are and what policy covers them.
For procurement and risk: ask what evidence exists that ads do not shape answers, beyond the vendor's description.
A wall between ads and answers is a design choice. Buyers and users should ask to see it tested.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: OpenAI.
