The first detailed study of advertising inside ChatGPT has found that lower-income users were more likely to be exposed to ads, while consumer goods dominated the emerging advertising ecosystem, offering an early glimpse into how generative AI could reshape digital marketing—and potentially reproduce some of the inequalities of traditional online advertising.
The study, The Beginning of ChatGPT Ads, published on August 5, describes itself as the “first empirical study of advertising content being rolled out in the user-facing online interfaces of large language models (LLMs).”
Researchers Emma Lurie, Ro Encarnación, Sorelle A. Friedler and Danaë Metaxa conducted what they describe as a “sock puppet audit” of ChatGPT’s advertising system, creating 91 simulated accounts designed to represent users from different income and racial/ethnic groups.
Across the experiment, the researchers collected and analysed 3,602 advertisements from 191 unique advertisers, generated through 335 prompts and more than 127,000 conversations.
Their central finding was striking: lower-income accounts were more likely to receive advertisements, irrespective of race.
The researchers found that advertising exposure decreased as the income associated with an account increased. Their analysis found an odds ratio of 0.98 for every $1,000 increase in median household income.
The result does not establish that OpenAI deliberately targets lower-income consumers with advertising. Rather, it identifies a pattern in the researchers’ simulated environment that they say warrants further investigation as AI advertising develops.
The study found no statistically significant relationship between race and ad delivery. The authors caution, however, that their sample was not sufficiently large to draw definitive conclusions about racial differences.
That makes the income finding particularly important because advertising has historically relied on systems that divide audiences according to perceived value, purchasing power and commercial intent.
AI assistants could introduce a new version of that logic—but with a crucial difference.
Instead of simply observing what consumers click on, conversational systems can potentially infer what they are thinking about.
AI becomes the next advertising frontier
The researchers argue that large language models are becoming more than information-retrieval tools.
They describe LLM chatbots as “the next frontier for online advertising”, pointing to the combination of conversational interaction, personalisation and information disclosed by users.
That combination gives conversational AI “considerable persuasive potential,” the researchers write.
It is this characteristic that could distinguish AI advertising from search and social advertising.
A search engine may know that someone is looking for a hotel. A social network may know that someone has shown an interest in travel. A chatbot, by contrast, may know that the person is planning a family holiday, has a particular budget, dislikes long flights and is trying to decide between two destinations.
The advertising opportunity is therefore not simply to reach a consumer.
It is to enter the consumer’s decision-making process.
That raises a fundamental question for the industry: when an AI assistant recommends a product, how clearly can the consumer distinguish between useful advice and commercial persuasion?
Consumer goods dominate the early AI advertising market
Despite the futuristic possibilities, the advertising ecosystem documented by the study looks surprisingly familiar.
The researchers found that advertisements “skewed heavily towards consumer goods.”
Their analysis shows that retail and information-sector advertisers together accounted for roughly 57% of the advertisements collected.
The researchers also found that ads generally promoted advertisers rather than individual products and were “clearly separated from the LLM’s response text.”
That separation is significant.
At this early stage, ChatGPT advertising does not appear to be seamlessly embedded inside the model’s answers. Instead, advertisements remain recognisable as advertising, maintaining a distinction between sponsored content and the AI-generated response.
But the researchers suggest this could change.
They describe their findings as representing the “first phase of ChatGPT ads” and anticipate that advertising will evolve “as ads continue being integrated into LLM chat interfaces.”
The implication is that today’s relatively simple formats may be only the beginning.
Why conversational advertising is different
Traditional online advertising has generally been built around attention and behavioural signals.
Consumers search, browse, scroll and click. Advertising systems use those signals to predict what products or services might interest them.
Generative AI introduces a different source of information: conversation.
Users routinely tell chatbots things they might never explicitly enter into a search engine. They explain financial constraints, family circumstances, preferences, fears, ambitions and plans.
The researchers point out that users increasingly rely on LLMs as “an information intermediary.”
That intermediary role could make advertising dramatically more persuasive.
If a user asks an AI which laptop to buy, which insurance policy to choose or where to spend a holiday, the AI is no longer simply displaying advertising alongside information. It is potentially occupying the position traditionally held by a trusted adviser.
And that is precisely why advertising inside AI raises concerns that are different from those associated with a banner or social-media advertisement.
Who gets advertised to?
The income finding becomes more consequential in this context.
The researchers’ simulated accounts were designed around neighbourhood-level demographic characteristics, allowing the study to examine whether advertising exposure varied according to income and race.
The researchers found that “lower-income accounts were more likely to receive ads” and that this pattern persisted across racial groups.
They stress that the study cannot explain why this happened.
Possible explanations could include differences in the types of conversations generated by the simulated accounts, differences in advertiser targeting, differences in commercial value assigned to particular audiences, or other characteristics of the advertising system.
The study therefore should not be read as evidence of intentional discriminatory targeting.
But it establishes something arguably just as important at this stage: advertising distribution in AI systems is already producing measurable demographic patterns.
As AI advertising becomes more sophisticated, those patterns could become increasingly consequential.
The business model looks familiar
The researchers place ChatGPT advertising within the much longer history of internet monetisation.
They recall a famous exchange from Facebook’s early years, when Mark Zuckerberg was asked how the company could make money if users did not pay.
His answer was:
“Senator, we run ads.”
The researchers argue that despite the dramatic technological transformation represented by generative AI, “the internet’s business model has not” fundamentally changed.
That observation highlights the tension at the centre of AI’s commercial future.
Running large language models is enormously expensive. Advertising offers AI companies a familiar way to subsidise free access.
But the nature of the product is different.
Facebook was designed around social interaction. Google was designed around search. ChatGPT is designed around conversation.
And conversation creates a different relationship between user and platform.
The trust problem
The researchers warn that the combination of conversational interaction and personal information gives LLM advertising unusual power.
Users may regard an AI assistant not simply as a platform but as an adviser.
That means the commercialisation of the conversation could affect the very trust that makes the technology useful.
The researchers describe their findings as an empirical baseline for “ChatGPT’s nascent advertising infrastructure.”
The word “nascent” is important.
The researchers are not presenting a mature advertising ecosystem. They are documenting its earliest stage.
The formats may change. Targeting may become more sophisticated. Advertisers may learn how to create campaigns specifically designed for conversational environments. AI systems may eventually understand not only what consumers want but when they are most receptive to a commercial suggestion.
That could transform advertising.
But it could also create a new regulatory and ethical frontier.
From search advertising to conversational persuasion
The central significance of the study is therefore not simply that ChatGPT is showing ads.
Advertising has already appeared on almost every major digital platform.
The significance is where those ads are appearing.
A chatbot can sit between a consumer and a decision.
It can help someone choose a product, compare services, plan a purchase or solve a problem. If advertising becomes part of that interaction, the boundary between information and persuasion becomes much harder to define.
The researchers describe LLMs as having “considerable persuasive potential.”
For advertisers, that may be the attraction.
For consumers, it may be the warning.
The study offers only an early snapshot of what is likely to become a much larger advertising market. But its findings already raise questions that the industry will have to confront as AI becomes a more important gateway to information and commerce.
Who receives the ads?
Why do they receive them?
What determines the advertiser they see?
And, most importantly, when an AI tells a consumer what to buy, will that recommendation still be trusted if an advertiser helped shape the conversation?
Those questions may ultimately determine whether AI becomes advertising’s most powerful new medium—or the place where consumers finally demand a different bargain.
