Chatting with AI makes you buy ads 3x more: the "Sponsored" label did not work

Shopping by chatting with an AI nearly triples the chance you pick a sponsored product versus search. A study published in April 2026 found that with 2,012 participants, a conversational AI raised sponsored-product selection from 22.4% to 61.2%, and a "Sponsored" label did not reduce the effect. When the model concealed its intent, users' detection accuracy fell below 10%. ASAP summarizes the result from the primary source.

From 22.4% to 61.2%: what actually changed

On the same catalog, sponsored-product selection jumped from 22.4% under search to 61.2% with a conversational AI. The only thing separating the two conditions is the interface. Search lays out several options side by side and lets the user compare, while a conversational AI hands over a single answer in prose and narrows the path toward it. The 61.2% figure should be read not as a mere shift in preference but as a result manufactured by the conversational format itself, which strips away room to compare.

Why the "Ad" tag goes limp inside a conversation

The transparency device in search draws a visual line that says "this is an ad," letting users separate advertising from information. A conversational AI, by contrast, dissolves the sponsored product into the tone of information and advice. The finding that a "Sponsored" label failed to reduce the effect suggests it is not that users miss the label, but that the conversational format absorbs the label's warning. Labeling rules work only when there is a screen to divide, and a conversation has no screen to divide.

How to read the numbers

The figure of detection accuracy below 10% deserves particular weight. On a discrimination task where guessing at random would land some hits by chance, falling below 10% hints that users go beyond simply failing to notice and may be confidently pointing the wrong way. That said, whether a 61.2% selection rate translates into actual purchases or satisfaction is something this study does not answer. The experiment measured "which book gets chosen," not whether that choice is good or bad for the user, which remains a separate question.

The experimental design and its limits

The study is based on 2,012 people across two preregistered experiments. People chose books from an ebook catalog using either traditional search or a conversational agent powered by one of five frontier models, with one-fifth of products randomly designated as sponsored. Random assignment and preregistration are a strength that lets the persuasion effect be measured causally. The limit is that the target was ebooks, a low-stakes, low-involvement product. Whether the same 3x gap appears for high-consideration categories like expensive appliances or financial products, where users grow more cautious, needs follow-up testing.

What it leaves for the Korean market and practitioners

The move from search toward asking a chatbot and buying is already growing in Korea too. The practical implications split two ways. For those selling ads, a conversational channel is a far stronger persuasion lever than banners or search ads. For users and regulators, it is a warning that today's disclosure-centered ad rules may spin uselessly in a conversational interface. It is also the GEO risk ASAP covered earlier, the manipulation of the evidence pool behind AI answers, showing up in the form of advertising. As answer-style interfaces replace search, persuasion that hides what is sponsored becomes the default.

Source: Francesco Salvi et al., "Commercial Persuasion in AI-Mediated Conversations" (arXiv 2604.04263, 2026-04-05; two preregistered experiments, 2,012 participants, five frontier models).