Ask an AI chatbot where to buy a product online and the answer can include a fake store or a counterfeit seller. That is the finding of a study by Podqi, an intellectual property protection firm, which tested ChatGPT, Google’s Gemini and Perplexity with the kind of shopping questions people actually type. About 9 percent of the 6,795 answers included a risky site, and in most cases the chatbot recommended it as a place to buy.
The study was shared exclusively with The Verge, which reported it on November 12. Podqi wrote its own prompts and ran them through the logged-out versions of the three services in two rounds during September, using data from AI search tracker Gauge. The questions covered 151 brands and asked where to buy a brand and whether it shipped abroad through regional sites. The figures reflect answers to those test prompts, not real shopping trips.
How the Chatbots Compared
Perplexity surfaced risky sites in 15 percent of its answers, followed by ChatGPT at 8 percent and Gemini at 4 percent. Podqi counted a site whether the chatbot recommended it, mentioned it without comment, or warned shoppers against using it, so the headline rates describe how often a risky site appeared, not how often a shopper was actively pushed towards one.
Even so, the split between recommending and warning is the most worrying part of the study. Across all three services, 57 percent of answers that surfaced a risky site recommended it as a place to buy, 34 percent mentioned it without comment, and only 9 percent warned shoppers away. Podqi said risky sites were recommended about 6.3 times more often than they were warned against. ChatGPT gave warnings most often, flagging the risk in 24 percent of such answers, but even that means three quarters of the time it showed a risky site without a warning.
Podqi reviewed 4,437 domains that appeared in the answers and classified 718 as risky, including 351 fake sites and 367 inactive ones. A fake site might impersonate a real retailer or sell counterfeit goods. An inactive domain no longer works but could be bought and reused by a fraudster, and an old brand domain re-registered by someone else is the same trap: the address looks familiar, but the shop behind it has nothing to do with the brand.
Regional Store Questions Carried the Highest Risk
Not all questions were equally dangerous. Questions about international shipping and regional storefronts surfaced risky sites in 16 percent of answers, the highest rate among the five questions tested. Many brands do not have a local store in every country, so a chatbot searching for one can land on a lookalike site instead. In one test, Perplexity identified the wrong website as Stussy’s official store and recommended it as a source of authentic products. Smaller brands were more exposed: questions about larger brands surfaced risky sites about half as often.
The timing makes the findings hard to ignore. The study lands just before the holiday shopping season, when bargain hunters turn to unfamiliar retailers. Black Friday falls on November 27 and Cyber Monday on November 30. A confident-looking chatbot answer can be the first step towards a payment page that should never have been trusted.
What the Companies Said
The three companies responded very differently. Google spokesperson Jessica Levis Boord told The Verge that Gemini can make mistakes and that Google displays source links so users can check its answers, and Google was pleased that Gemini had the lowest rate of the three. OpenAI spokesperson Taya Christianson said the company is improving its safeguards, without disputing the findings. Perplexity spokesperson Jesse Dwyer pushed back, pointing to the company’s research benchmarks and saying the service does not claim perfect accuracy.
Those answers are worth reading alongside the caveats in the study itself. Podqi sells brand protection, so it benefits when impersonation looks like a serious threat. It wrote its own prompts rather than studying real conversations, and Madhappy, one of the 151 brands tested, is a Podqi client. The study also did not track whether anyone clicked a risky link, bought anything or lost money. The percentages are a snapshot of a test, not a universal rate. None of that erases the core finding, but the numbers should be read carefully.
Podqi chief executive Trevor West advised shoppers to verify a website before entering personal information or payment details, and to start from the brand’s main website for regional stores rather than trusting a search result or a chatbot link. The deeper lesson is about how people shop now: chatbots are becoming shopping assistants, and their confident tone makes a link feel like a recommendation even when it is only a guess. Until the services warn about risky sites as often as they recommend them, the safest habit is the oldest one: go to the brand’s own website, or a retailer you already know, and treat any unfamiliar store a chatbot suggests as a lead to check.
