Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Advertising

ChatGPT Will Show Ads While It Generates Your Image. Free Users Who Switch to Ads-Free Lose Image Generation

October 7, 2026 6 min read

On Monday, October 5, OpenAI announced a new visual ad format for ChatGPT. The first test runs during image generation, for users on the Free and Go plans, starting later this month in the US with an initial group of advertisers. OpenAI says the ads will be clearly labeled and will "remain separate from the image being created." The same post expands measurement for advertisers and announces brand suitability pilots with DoubleVerify and Integral Ad Science. OpenAI puts ChatGPT's reach at 1.2 billion people a week.

TechCrunch summed it up as visual ads that appear alongside image generation results. I build products on top of language models, so I spent this morning reading every ads page OpenAI has published since January, plus the help center and the ad policy. The format is the headline. The documentation is the story.

Nine months of expansion

OpenAI published its ads principles on January 16, answer independence and conversation privacy among them. The US test started on February 9, for logged-in adults on Free and Go. Go launched in the US at $8 a month, so from day one a paid plan carried ads. Only Plus, Pro, Business, Enterprise and Edu are ad-free.

Then came eight more markets, including the UK, Mexico, Brazil, Japan and South Korea, and on August 18 OpenAI announced 31 European countries, France included, starting the following week. On the advertiser side, bidding expanded beyond CPM and CPC to conversion optimization, with geo-targeting and custom audiences. By August, OpenAI said tens of thousands of marketers had advertised.

The ad policy has a changelog: version 1.0 in March, version 1.6 in September. Two of the revisions concern my field. In April, medical, legal and financial advice contexts were "no longer categorically blocked from ads by default." In August, ads for legal services became permitted in the US, for advertisers licensed in the jurisdiction where the ad is shown. I build a LegalTech product. A US user on the free plan who asks ChatGPT a legal question can now be eligible for a legal services ad under the answer, as long as the conversation itself is not classified as sensitive.

The wall only goes one way

OpenAI's central promise is answer independence: ads do not influence answers. The help center describes the mechanism: "Ads run on separate systems from our chat model." By default, ChatGPT cannot even see the ads shown to you. You have to tap "Ask ChatGPT" on an ad to share it with the model.

Now look at the other direction. To pick an ad, the system considers "the context and intent of your current conversation", plus general location and language. With ad personalization on, it can also use past chats, memory and the way you reacted to previous ads. Turn personalization off and you still see ads "based on the context of your current chat thread." Personalized ads are not initially available in the EEA or Switzerland, so here in France the current thread is the main signal.

So the model does not read the ads, but the ad system reads the conversation. Advertisers never receive chats, memories, names, emails, precise location or IP addresses, only aggregated performance. That is a real protection. It is also a narrower promise than "nobody reads your conversation for commercial purposes."

Even age is inferred from behaviour: to keep ads away from under-18s, the age prediction model considers "stated age, how long the account has existed, typical times of use, usage patterns over time, and the general topics a person discusses."

Measurement brings data in, not out

The October 5 measurement post shows the plumbing. Advertisers send conversion data through the OpenAI Pixel and a Conversions API, with integrations for Hightouch, Tealium and LiveRamp and attribution partners such as AppsFlyer and Adjust. Reports can include a one-day view-through window: a purchase after an impression, without a click, still counts. OpenAI says 52.7 percent of eligible one-day view-through conversions happened within an hour of "the matched ad impression."

Read the privacy promise and the measurement post together and the shape is clear. Conversations do not flow out to advertisers. Conversion events flow in from advertisers and get matched to impressions inside OpenAI. That is how every large ad platform works. It is new for the tool where people paste their contracts, their drafts and their business plans.

What changes for people who build on these models

None of the ads pages I read mention the API. Everything they describe is a ChatGPT surface: a unit below a response, or now during image generation, chosen by a separate system. If your product calls a model through an API and renders its own answer, what your user sees is still yours. Say so out loud, to clients and to your team.

Here is what I am changing on Monday.

First, an account audit. Anyone on my side who touches client material works in a business plan or through the API, never in a personal Free or Go account. With client documents, that was already the rule for confidentiality. It is now also the rule because a Free thread is an ad-targeting input. If your sales or support team drafts in free ChatGPT, check that first.

Second, a recommendation drift test. Answer independence is a policy, and I cannot audit it from outside. What I can audit is my own product. Wherever a model in my stack names a vendor, a tool or a provider, I keep a fixed set of commercial-intent prompts, run them on every model change, and log which names come out. If the names shift, I want to know before a client does.

Third, write down my own ads principle before I need the revenue. Our position at CBlindspot is simple: no third party pays to appear next to an answer. In writing, it becomes a commitment clients can hold us to, instead of a default nobody chose.

The business model question

The most honest page in the set explains how to avoid ads. On Free, you can switch to an ad-free mode, with fewer messages and no access to some tools, "like image generation or deep research." Image generation is one of the tools you give up to go ad-free on Free, and on October 5 it became the feature with its own visual ad format. That tells you which part of the free product OpenAI considers too expensive to give away unless someone else pays.

Anyone who sells AI features faces the same arithmetic at a smaller scale. Free inference is paid by someone: by usage caps, by a subscription, or by a third party. OpenAI now uses all three at once, plus a paid tier that still shows ads. For my side projects on cheap API tiers, caps and subscriptions are the honest options. For a B2B product that handles inventions before they are filed, a third party in the conversation is not an option at all.

If you sell an AI product and hoped to advertise it in ChatGPT, note one line added in September: OpenAI reserves the right to decline ads that conflict with its "business interests, or competitive position."

I build on Claude, so read this with that bias in mind. Anthropic wrote on February 4 that Claude "will remain ad-free", and argued that ad incentives, once introduced, "tend to expand over time." OpenAI's principles promise it will "always offer a way to not see ads in ChatGPT, including a paid tier that's ad-free." Both are policy pages, not contracts.

When the assistant your users talk to is also an ad surface, the line between the answer and the commercial layer becomes a product feature. If you build on these models, you get to decide where that line sits in your own product. Decide it on purpose, and write it down.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk