OpenAI has published a detailed look at how ChatGPT ads are chosen, and the short version is that a language model works from a summary of your conversation to decide which ad might fit. On October 21, 2026, the company also said ChatGPT ads are expanding to 26 more countries, for 89 in total.
If you use the Free or Go plan, ads can appear below an answer, so it is fair to want to know what that system reads. The post explains a lot, but it leaves out the settings most people would go looking for.
TL;DR: OpenAI says ChatGPT ads are picked in a separate service that turns your chat into short summaries, searches an ad index, and runs an auction. The ad model does not pick the winner or set prices. Ads are labeled as sponsored, and showing no ad is a normal outcome. The post names no user settings, so what you can switch off is still unclear.
How ChatGPT ads get picked
OpenAI describes a pipeline with a language model it calls the orchestrator at the front. It turns the conversation into brief summaries and rewritten search queries, which become semantic embeddings used to search several ad indices.
After that, the steps look like familiar ad tech with a model bolted onto the start. The table below follows the order OpenAI gives in its engineering post.
| Stage | What OpenAI says happens |
|---|---|
| Summarize | The orchestrator condenses the conversation into short summaries and rewritten queries |
| Retrieve | Nearest neighbor search pulls a broad candidate pool, including a product feed index that accounts for reviews |
| Filter | Candidates drop out on availability, locale, policy, budget, pacing, placement and per advertiser duplicates |
| Score | Survivors are scored on semantic relevance, predicted engagement and historical performance |
| Auction | Admission thresholds, pacing, quality adjusted ranking, pricing and final selection |
| Final check | The orchestrator can block an ad that feels inappropriate in context |
The orchestrator nominates candidates to the auction, but OpenAI says it does not pick the winner or set prices. That separation is worth noticing, because it limits how much one model’s judgment can steer what you see.
What matters more to most readers is what goes into that first summary step.
What data feeds the ad decision
The system draws on the current conversation and its history, including sessions that persist across turns and across multiple conversations. OpenAI’s own example is a grocery delivery ad shown while someone plans meals for a marathon.
If you choose to personalize ads, it can also use nonsensitive information from your past conversations. Clicks, hides and surveys count as signals too, along with feedback from advertisers through Ads Manager.
Missing from the post is any list of topics that count as sensitive, or any rule for minors. OpenAI says only that integrity models decide when a conversation is appropriate for an ad. Minors are a topic OpenAI has already faced questions on, as in DigitBin’s report on ChatGPT for Teens parent alerts, so that gap is the one to watch.
Where the ad lands, and how it is kept apart from the answer, is spelled out more clearly.
Where ads appear and how they stay separate
Ads appear below ChatGPT’s response, only on eligible turns, and they are clearly labeled as sponsored. A model-generated line explains how the ad relates to the conversation, and OpenAI keeps that line apart from the advertiser’s own creative so ChatGPT’s voice stays neutral.
Ad selection runs in a separate service with its own models and code. OpenAI says this is why ads do not influence organic answers, and why the response carries on normally if the ads system is slow or down. Ads also have to arrive before the chat response is complete.
Returning no ad is a normal outcome. So a reply with nothing under it is the system working as designed, not a glitch.
On plans, OpenAI’s expansion announcement says ads appear on Free and Go, while Plus, Pro and Enterprise stay ad free. That makes a paid tier the only ad free route the company describes.
The business numbers explain why OpenAI is publishing all of this now.
Why OpenAI is spelling this out now
OpenAI says ChatGPT ads passed $1 billion in annualized revenue run rate in under 200 days. More than 100,000 advertisers have run campaigns since ads launched just over eight months ago, and OpenAI says more than 1.2 billion people use ChatGPT weekly.
The 26 new countries include Argentina, Chile, Colombia, Peru, Nigeria, South Africa, Serbia, Ukraine, Cambodia, Bangladesh and Pakistan, and OpenAI does not list the rest. Advertisers reach them through technology partners or self service in Ads Manager.
The post also promises more formats, objectives, buying options and measurement tools. One of those is already in view, since DigitBin covered OpenAI’s plan to trial image generation ads in the US.
What to check in your own account
OpenAI says personalization is a choice you make, and its page footer carries a “Your privacy choices” link, but the post never says where that choice lives inside the app. Until it does, assume the free tiers read a summary of what you type.
If you often ask ChatGPT about money, health or family, that is worth keeping in mind before you rely on the ad separation alone. It is a reasonable design on paper, and the proof will be in how clearly the controls show up.
