AI chatbots pointed shoppers to fake sites in up to 15% of answers

Chatbots from Google, OpenAI and Perplexity are recommending fake or risky shopping websites to users, according to a study by IP protection firm Podqi shared with The Verge. The firm ran 6,795 shopping-related queries about 151 brands, including Lululemon, Birkenstock, Stüssy, Madhappy and Le Creuset, through Gemini, ChatGPT and Perplexity over two testing periods in September.

Podqi defined a risky site as one impersonating a real retailer, selling knockoffs, occupying a brand’s old reregistered domain, or sitting dormant and available for someone else to claim. Perplexity performed worst, citing or displaying a risky site in 15 percent of its answers. ChatGPT landed in the middle at 8 percent, while Gemini showed risky sites least often, at 4 percent. ChatGPT was also the most likely to flag the risk to users, warning about a risky site in 24 percent of the cases where one appeared, according to Podqi.

Smaller brands fared worse across all three tools. Podqi found that questions about larger companies returned risky sites roughly half as often as questions about smaller ones. Prompts asking about regional brand websites were especially prone to triggering bad results, and the chatbots sometimes invented regional sites for brands that don’t operate any, Podqi found. In one example cited in the study, Perplexity misidentified a site as Stüssy’s official store and described it as the best place to find guaranteed authentic items.

Asked about the findings, Google spokesperson Jessica Levis Boord said the company takes information quality seriously and is glad Gemini performed best among the three, while noting that Gemini is upfront about being able to make mistakes and displays source links so people can verify answers. OpenAI spokesperson Taya Christianson said the company wants ChatGPT to connect people with reliable places to shop and is working to make deceptive sites less likely to appear in results. Perplexity spokesperson Jesse Dwyer said the company has focused on research accuracy for four years and does not claim to be perfect.

What this means

The findings point to a gap between how AI companies market chatbots, as useful tools for researching and comparing products, and how reliably those tools can verify where a link actually leads. Podqi’s data suggests the risk is uneven: shoppers asking about well-known brands get safer results than those asking about smaller labels or regional storefronts, which is exactly the kind of query a holiday shopper chasing a deal might type.

It is not clear from the study how often users who encountered a risky site actually clicked through or entered payment information, since Podqi’s test used crafted prompts rather than real user sessions. Podqi cofounder Trevor West says shoppers should still verify a source independently before entering personal or card details, even when a chatbot presents it as a direct answer.