Study: Risky web stores appeared in 9 percent of 6,795 chatbot answers about shopping
A study from IP protection firm Podqi, shared exclusively with The Verge, shows that ChatGPT, Gemini and Perplexity can all surface suspected fake, imitated or inactive web stores in answers to shopping questions. But the percentages measure exposure — a risky site appearing in an answer — not confirmed recommendations of scam sites or measured fraud. The numbers therefore require a nuanced reading, especially right before the holiday shopping season.
What the study did
Podqi tested the three chatbots with five shopping questions about 151 brands, in two test waves in September 2026, against logged-out consumer services. The answers were collected through Gauge, a company that tracks AI responses. In total, 6,795 answers were analyzed, according to remio.ai's review of the study.
The researchers assessed 4,437 domains and classified 718 of them as risky: 351 suspected fake websites and 367 inactive domains. The category covered imitation and counterfeit sites, sellers of copy products, old brand domains that have been re-registered, and dead domains.
The study was published on November 12, 2026 via The Verge — right before holiday shopping, when a growing number of consumers use AI chatbots to find products and stores.
The numbers — and how to read them
Perplexity fared worst: risky websites appeared in 15 percent of its answers. ChatGPT was in the middle at 8 percent, and Gemini did best at 4 percent. Aggregated, 9 percent of the 6,795 answers contained a site Podqi classified as risky, according to remio.ai.
Here is the essential point about the method: Podqi does not only count when a chatbot recommends a risky website, but also when it is merely mentioned, and when the chatbot warns the user against it. The correct reading is thus "a risky website appeared in 15 percent of the answers," not "the chatbot recommended a fake store in 15 percent of the answers."
Within the share of answers that contained risky sites, the findings break down as follows: 57 percent recommended them, 34 percent mentioned them without comment, and 9 percent warned the user. ChatGPT was best at warnings, cautioning users in 24 percent of the cases where risky sites appeared — meaning that in roughly three quarters of cases, the sites were listed without any form of comment.
In other words, the study measures exposure, not fraud. None of the sources quantify actual consumer losses or measure fraud attempts resulting from the answers.
Concrete examples and patterns
The clearest example in the reporting is the clothing brand Stüssy: in one conversation, Perplexity named the wrong site as Stüssy's official web store and described it as "the best starting point for guaranteed authentic goods" — translated from The Verge's rendering of the test answer.
Two patterns stand out. First, small brands are particularly exposed: questions about larger companies produced risky websites roughly half as often, meaning roughly double the risk for the smaller ones. That makes intuitive sense — larger brands likely have better domain enforcement and more takedown data for the chatbots to draw on — but Podqi does not offer a mechanism in the reporting that explains the connection. Second, region-specific questions about websites triggered risky results more often, suggesting that the chatbots struggle most precisely where users have the hardest time verifying the answers themselves.
The companies' responses
Google, which came out best in the test, had spokesperson Jessica Levis Boord write in an email to The Verge: "We take information quality seriously, and are glad to see Gemini performed best among the products tested." Google also points out that Gemini shows source links so users can verify the information themselves — a company statement about its own performance.
OpenAI offered, in the same channel, a goal-oriented statement: "Our goal is for ChatGPT to connect people with trusted places to shop," wrote spokesperson Taya Christianson, stating that the company is improving safety mechanisms against deceptive websites in shopping results.
Perplexity contests the framing. Spokesperson Jesse Dwyer wrote in an email: "Perplexity is consistently the most accurate AI on credible research benchmarks, and has been for four years. We never claim to be 100 percent accurate, but we do claim to be the only company uncompromisingly focused on it." It is worth noting that Perplexity does not dispute the findings themselves, but rather the focus and the definition of "risky."
Open questions and limitations
Several caveats limit how far the findings can be stretched:
- The study is not publicly available. All numbers in this story come from secondary reporting — The Verge, which received the study exclusively, as well as reviews from remio.ai and AI研究所. We have not read the underlying report ourselves.
- Exposure, not fraud. The study shows that risky sites appeared in answers. It does not measure whether anyone was actually deceived, shopped on fake sites, or lost money.
- A limited setup. The test used five fixed questions about 151 brands, in September 2026, in English, in a US context, without logging in. AI研究所 notes that it is unknown whether the same results would emerge for questions about Japanese stores in Japanese — and the same applies to other languages and regions, including Norway. Logged-in setups and agentic shopping features were not tested either.
- Unclear details in the numbers. The Verge describes "a handful" of questions while remio.ai says five; such differences are compatible, but show that the details of the method cannot be fully verified from public reporting.
Why it matters
The finding points to a concrete weak spot in AI-assisted shopping: when chatbots name websites with confidence — as in the Stüssy case — the user has little reason to suspect that the domain is fake or dead. The weakness is greatest for small brands and region-specific questions, precisely where consumers have the least knowledge of their own to fall back on.
For brands, the point is that domain monitoring and IP protection are no longer just about removing scam sites from search engines — it is also about how they appear in AI answers. And for consumers, in the middle of holiday shopping, the most practical advice is the one implied by the findings themselves: check that the site you actually shop on is the official channel — not just the one the chatbot pointed to.

