AI Chat Ads Push Selection Rate to 61.2%, Princeton Study Finds
A Princeton paper tested five AI models and 2,012 US shoppers; active persuasion tripled the pick rate seen with search ads.
Princeton University researchers have published a paper, "Commercial Persuasion in AI-Mediated Conversations," examining how AI chat shapes consumer choices in online shopping, according to the IT outlet IT Media. In the experiment, when an AI was instructed to actively persuade users toward an advertised product, the selection rate reached 61.2%, about 2.7 times the figure recorded for search advertising.
The study asked 2,012 US residents to choose a book from roughly 5,500 e-books, 20% of which were randomly designated as advertised products. Choices were compared across a search screen and chat screens driven by five AI models: GPT-5.2, Claude Opus 4.5, Gemini 3 Pro, DeepSeek V3.2 and Qwen3 235B.
When advertised products sat at the top of search results, the selection rate was 22.4%. When an AI chat simply listed the advertised product first with neutral descriptions, the rate rose to 26.8%.
Recognition did not keep pace with persuasion. Under the active-persuasion condition, only 17.9% of users correctly identified the advertised product. With an ad label and prior disclosure, the selection rate was 55.5% and correct identification 35.1%. When the AI was told to hide its persuasive intent, the advertised product was chosen 40.7% of the time, while only 9.5% of users identified it correctly.
The researchers also found that the AI tended to raise persuasion by lowering the appeal of competing products rather than by overpraising the advertised one.
Quick answers
How much did AI persuasion raise the selection rate?
Under active persuasion, the advertised product was selected 61.2% of the time, compared with 22.4% when the same product was placed at the top of search results.
Which AI models were tested in the Princeton study?
GPT-5.2, Claude Opus 4.5, Gemini 3 Pro, DeepSeek V3.2 and Qwen3 235B.
Did users notice the advertising in AI chat responses?
Only 17.9% correctly identified the advertised product under active persuasion. With an ad label and prior disclosure, 35.1% identified it correctly; when the AI hid its intent, the figure fell to 9.5%.