Study finds risky sites in 9% of AI shopping answers

About nine in every 100 answers to shopping questions brought up a suspected imitation store, a site thought to sell counterfeits or an inactive address in a test of ChatGPT, Gemini and Perplexity. Brand protection company Podqi analyzed 6,795 answers collected in September. The Verge reported the findings and responses from the AI providers on November 12.

Among the answers in which a risky address appeared, 57 percent presented it as a place to buy something. Another 34 percent mentioned the site without comment, and 9 percent warned shoppers away. Podqi counted each type separately. A warning can surface the same address as a recommendation while giving the reader very different information about the seller.

What the researchers counted

The researchers selected 151 brands, including apparel and home goods companies, and put five kinds of shopping questions to each service. They ran two rounds in September and collected 2,265 answers from each service. The queries were written for the study. Its percentages describe the answers to those questions. The researchers did not sample live customer chats or completed orders.

Podqi classified 718 of the 4,437 distinct domains it reviewed as risky. It identified 351 as suspected of impersonation or offering counterfeit goods and 367 as inactive. The same address could show up more than once. The domain tally measures the number of different addresses, while the answer tally measures how often the systems exposed readers to one.

Perplexity mentioned risky sites in 15 percent of its answers, followed by ChatGPT at 8 percent and Gemini at 4 percent. ChatGPT warned about the particular site in 24 percent of its answers that surfaced a risky address. Questions about regional brand sites produced the highest combined rate in the test, at 16 percent. The figures apply to this set of brands and questions. Podqi sells services that investigate brand impersonation and trademark misuse.

From a search result to the seller

OpenAI’s guide to shopping with ChatGPT Search says product cards can display images, details and links to sites where people can learn more or buy. Selecting an item can bring up other sellers and their prices. The company says changes to a merchant’s prices or shipping terms can take time to show up in ChatGPT, so the first price on the screen can differ from the seller’s current offer.

In March, OpenAI said it was making it possible for more merchants to supply product catalogs and promotions to ChatGPT. It also described a shift away from the initial version of its standalone checkout, letting sellers provide their own purchase experience. A shopper can discover a product in one place and make the purchase with a merchant elsewhere. Podqi examined the addresses that appeared in the discovery stage.

The U.S. Federal Trade Commission’s online shopping guidance treats the seller’s reputation, total cost including shipping, and return terms as separate checks. It also explains the limits of HTTPS. An encrypted connection protects data while it travels between a customer and a site. The address can still belong to an impersonator.

Clicks from AI to U.S. retail sites

Adobe reported on October 29 that U.S. retail visits from AI sources were up 71 percent year over year in September. Its measure counts a visit when a person clicks a link and reaches a retailer. Podqi counted addresses in answers, including warnings, whether or not anyone followed them. Adobe’s analysis measures traffic rather than the trustworthiness of the destinations.

Adobe forecast that AI-referred visits on Black Friday would rise 95 percent from a year earlier. The forecast came ahead of the shopping event. OpenAI told The Verge it was working to reduce the appearance of potentially deceptive sites in shopping results. Google said it shows links alongside Gemini responses so people can check their sources.