Chatbot Ads Turn Profitable, 72% Say They Are Uneasy

Advertising has moved inside the AI assistant. ChatGPT and Microsoft Copilot both carry ads today, and survey data published this month suggests most Americans are not comfortable with the data that makes those ads work.

In a Harris Poll survey of 2,066 US adults, 72% said they are concerned that AI chatbots will use their conversation data to target them with ads. Of those, 44% described themselves as very or extremely concerned and 28% as somewhat concerned. The survey was commissioned by eyeo, the company behind Adblock Plus, and published Aug. 17 in a report on advertising inside AI assistants.

AI has become a first stop for product research, and that is what makes conversation data commercially valuable.

Why the Ads Are Arriving Now

Among people who use AI in purchase decisions, GWI found that 33% use it for detailed product information, 31% for price comparisons and 30% to compare features. Those are buying signals stated in plain language while the purchase is still being decided.

On the cost side, Goldman Sachs estimates $7.6 trillion in capital will be required over the next five years for compute, data centers and power. Advertising is one of the few revenue models that scales against a number that size. Reuters reported in March that OpenAI’s US ad pilot passed $100 million in annualized revenue six weeks after launch.

Google is the holdout for now. The Gemini app carries no ads, though Alphabet said on its first-quarter earnings call that its AI Mode ad formats would transfer to the Gemini app.

What the Ads Actually Look Like

Ads appeared on ChatGPT in February 2026. The two live implementations differ in a way that matters.

• ChatGPT shows a plain box containing the advertiser’s own creative, placed after the answer and marked with an “ad” label.
• Copilot writes its own line of copy in the assistant’s voice, tied to what the user just asked, and places that sentence above a sponsored box carrying a Microsoft Advertising label.

In both products, the ad was matched to the prompt and the surrounding context rather than to the answer itself. In some sessions the assistant recommended a competitor of the brand paying for the adjacent placement.

Those observations come from test sessions run by eyeo in August on free plans in Chrome, using fresh accounts with no conversation history. The company describes them as a small, non-representative set and a record of what it saw rather than how often it happens. NervNow did not independently replicate the tests.

What Users Say When Asked Openly

A separate open-ended survey of 150 US adults, also commissioned by eyeo and run by Pollfish, asked what people make of ads appearing inside AI assistants. Responses were coded into non-exclusive themes. Annoyance or intrusion came up in 17% of answers. Resignation, along the lines of this being inevitable, accounted for 14%. Another 9% accepted ads as the price of a free product. And 7% said ads would make them trust the answers less.

Cornelius Witt, director of global public affairs at Adblock Plus, said the report was intended to make the mechanics visible. AI assistants have become central to how people find information, he said, and yet little is publicly known about the ads inside them: how they work, how many there are, how they are selected, and what users think of them.

“We want to open this black box.” Cornelius Witt, director of global public affairs, Adblock Plus, whose employer sells software that blocks these ads

The Harder Problem Is Coming

Two academics interviewed for the report pointed to the same next step: advertising written into the answer rather than placed beside it.

Florian Schaub of the University of Michigan School of Information said the shift is structural. “Injecting advertising into chatbot responses fundamentally changes how ad delivery works,” he said. On the web, advertisers prepare assets in advance and targeting selects which one a user sees. In AI chatbots, targeting works in similar ways, but the prepared assets are no longer needed. By Schaub’s account, the system needs only a keyword and a product matching prompt, after which it can generate the copy itself, shaped by the user’s characteristics and the chat context.

“I’m really curious to see how far these systems will go in using their own inference capabilities to target ads next.” Florian Schaub, University of Michigan School of Information

Ines Zelch of Friedrich Schiller University Jena is researching that scenario directly. Her group has injected brand and product mentions into generated text as a form of product placement and trained models that detect it well, though so far only on artificially generated data rather than real output. An early user study found that many people do not expect advertising inside generated text and often fail to notice it, which leaves them open to influence. Zelch also flagged a subtler version: shifting a user’s preferences over weeks or months without naming a product at all.

What This Means if You Buy or Sell Attention

• New inventory, no standard measurement. Ads inside AI answers are live and generating revenue, but there is no shared framework yet for viewability, attribution or frequency.
• Adjacency risk is real. If placement is keyed to the prompt rather than the answer, a paid slot can sit directly beneath a recommendation for a rival.
• Conversation data is the targeting signal. That is the specific thing 72% of surveyed US adults said they are concerned about, and it is a privacy and disclosure question before it is a media question.
• Disclosure gets harder from here. A labeled box is easy to identify. A product mention generated inside a sentence is not, for users or for the tools built to filter it.

For two decades, web advertising ran on the premise that an ad was visibly an ad. Both live implementations still label their placements. The research suggests the next generation may not be separable from the answer at all.