Podqi Finds Risky Stores in ChatGPT, Gemini and Perplexity Answers
Podqi checked which stores ChatGPT, Gemini and Perplexity show people looking to buy products. Risky URLs appeared in answers from all three services.
In September 2026, Podqi checked which stores ChatGPT, Gemini and Perplexity show people searching for products. According to Habr, risky URLs appeared in 15% of Perplexity’s answers, 8% of ChatGPT’s and 4% of Gemini’s. Those figures do not represent the share of answers recommending a purchase from an unsafe site: the researchers also counted neutral mentions and warnings.
How the test worked
Podqi, an intellectual property protection company, prepared questions about products from 151 brands. They ranged from small businesses to major manufacturers, including Lululemon, Birkenstock, Stüssy and Le Creuset. Madhappy, a Podqi client, was also in the sample. The researchers asked where to buy products, which stores to choose, and whether international shipping and regional sites were available.
The test had two stages. Podqi wrote the prompts, not real shoppers, and collected the answers using an AI search results tracking tool. In total, the team examined 6,795 answers. The figures apply to those prompts and to testing in September, so they cannot tell us what happens in every shopping conversation with a chatbot.
The risky category included sites impersonating official stores or suspected of selling counterfeits, former brand domains that had changed hands, and pages that no longer worked. The last group is not necessarily associated with fraud. But an unavailable domain could acquire a new owner. If that happens, a link the chatbot keeps showing shoppers will lead somewhere else.
Where the chatbots got it wrong
Answers about regional stores were especially prone to problems. The services named supposedly official sites for individual countries even when the brand had no such sites. For example, Perplexity mistook a third-party page for a Stüssy store and presented it as a place to buy genuine products. For shoppers, it is not just the accuracy of the brand name that matters: the error is in the URL they are directed to visit.
Smaller brands proved more vulnerable than larger ones: risky sites appeared roughly half as often in answers about products from major companies. But the overall percentages should not be read as a ranking of shopping safety. They combine recommendations to use a store, simple mentions of a URL and warnings. ChatGPT, for instance, warned about the risk more often than the others—in 24% of its answers containing such sites.
Google, OpenAI and Perplexity commented on the findings to The Verge. Google noted that Gemini showed risky sites less often than the other services tested and pointed to the source links alongside its answers. OpenAI said it was improving protections against potentially deceptive sites in shopping recommendations. Perplexity called accuracy a priority but stopped short of promising that its answers would be 100% reliable.
What this means for stores and shoppers
For a seller, the important point is that a chatbot can make a mistake before someone even reaches the store’s site. If a third-party URL is presented as official, accurate product and shipping information on the real site will not correct the shopper’s route. It is particularly useful to keep an eye on answers about smaller brands and regional sales: those were the parts of the test where researchers found problems most often.
A marketer or developer can look at the route to a store from a shopper’s perspective: ask about sellers, shipping and sites for different regions, then check the domains named in the answers. This kind of check will not provide a lasting assessment of a chatbot’s performance—the answers in the study depended on prepared prompts. But it can help reveal when a service presents an old domain or a third-party page as a brand’s site. Another blog post about AI search tools also explains why the mere presence of a link in an answer should not be taken as proof that it is trustworthy.
For shoppers, a store appearing in an AI answer is no substitute for checking its URL. Podqi CEO Trevor West advises checking where a recommendation came from before entering personal information or card details. A warning is helpful, but the absence of one does not mean the store has been verified.
My take: the URL matters more than a confident answer
The mistakes with regional sites stood out to me most. In e-commerce, this is not just an inaccuracy in a product description: someone asks where to go to buy something and gets a specific URL. If a chatbot confuses an official store with a third-party site, the error lands right in the path to placing an order. I would check answers to the kinds of questions shoppers ask from time to time, focusing on the domains, not just the brand names.
That said, I would not label services “safe” or “unsafe” based on one table of percentages. Podqi tested its own sample during a specific period, and the overall figures even included answers with warnings. For me, the takeaway is practical: AI is useful for finding products, but you cannot trust a store just because a chatbot names it confidently.
Sources
Where the news comes from. The text is a retelling in the author’s own words; the facts come from the source, the opinion is the author’s.
e-commerce full-stack developer
I build online stores and integrations, from catalogs to order processing. I like tools that take care of repetitive work, but before adopting them I check how they’ll be updated and debugged later.
All posts by the authorRelated articles
Four AI Search Tools Cited a Made-Up Concept Within a Week
Researchers published articles about a made-up concept and found that ChatGPT, Perplexity, Grok and Google AI cited them. The authors caution that a citation in an AI answer does not, by itself, verify a claim.
OpenAI’s API and Google Found Only Some of the Pages ChatGPT Cited
Google and the OpenAI Responses API found only about a quarter of the pages ChatGPT cited. Measuring visibility in AI answers means distinguishing between matching websites and matching URLs.
Yandex Service Outage Affects Marketplaces, Delivery and Payments
On the morning of October 11, users reported disruptions to Yandex services and other companies’ services. The problems affected websites, apps, checkout, payments and delivery.
Discussion 4
Nikolaj Trnka
I’m not convinced the 15%, 8% and 4% figures tell store owners much about purchase risk: Podqi counted neutral mentions and warnings alongside recommendations, and it wrote the prompts itself. The URL is worth checking, but before paying for monitoring or changing anything, I’d want to see how often a chatbot actually sends a shopper to an unsafe store and what that costs a business.
Rūta K.
One detail worth separating: the 15%, 8% and 4% figures are the share of answers with risky URLs, not the share that recommended buying from an unsafe store. Since neutral mentions and warnings were included too, your question about actual shopper referrals is still the key one for deciding whether a small shop should spend money on monitoring.
Aigerim S.
@nik_toman We’ve had customers land on lookalike stores after finding a product through search, and the cleanup was much more expensive than checking where the link actually goes before sending traffic. That’s not proof these chatbot figures translate into lost sales, obviously, but I’d start with a manual spot-check of the URLs they mention before paying for monitoring. The awkward bit is that one bad recommendation can cost trust even if it’s rare.
Giorgi M.
That’s the practical distinction: the article’s percentages don’t tell us how often shoppers actually get sent to a risky store, but a quick manual URL check is cheap and can catch lookalikes before they become a trust problem.