AI chatbots are sending shoppers to fake and dead stores at a worrying rate, according to a new study by brand-protection firm Podqi. Risky sites appeared in 15% of Perplexity's shopping answers, 8% of ChatGPT's and 4% of Gemini's — and in most cases the chatbot actively recommended the risky site rather than warning about it.
Podqi tested ChatGPT, Google Gemini and Perplexity with shopping questions across 151 brands, including Lululemon, Birkenstock, Stüssy, Madhappy and Le Creuset. The testing ran across two rounds in September 2026 and collected 6,795 answers in total. That scale makes this one of the more substantial independent tests of AI shopping answers to date, and the brands chosen were the kind of well-known names whose popularity makes them prime targets for impersonation.
Across all three chatbots, about 9% of answers exposed a shopper to a risky site. Of the 4,437 domains in the answers, 718 were flagged — 351 as fake and 367 as dead. The dead ones are a particular trap: expired brand domains re-registered by someone else, so a link that once led to a legitimate store now leads somewhere entirely different.
The breakdown by platform is stark: risky sites appeared in 15% of Perplexity's answers, 8% of ChatGPT's and 4% of Gemini's. Perplexity, which leans hardest on web citations and real-time search, surfaced risky sites at nearly four times Gemini's rate — a gap suggesting each system's retrieval and verification directly affects how often shoppers are sent somewhere dangerous.
The most alarming finding is not that risky sites appeared, but how: they were recommended about 6.3 times more often than warned against, and in 57% of answers containing one, the chatbot actively suggested it for shopping. The AI was not merely failing to filter out bad links; it was endorsing them with the same confident tone it uses for legitimate ones.
The worst single case involved Stüssy. Perplexity named an incorrect page as Stüssy's official website and described it as a source of authentic products. For a shopper following that recommendation, the result would be a purchase from a site with no connection to the brand — at best a counterfeit, at worst a straight scam that takes the money and sends nothing. The chatbot's confident framing is what makes this dangerous: users trust AI answers precisely because they sound authoritative.
All three companies responded to The Verge, which reported the study. Perplexity spokesperson Jesse Dwyer defended the company's accuracy-benchmark record. The responses show the platforms know this is now a public trust issue, not just a research curiosity — how they respond in product terms is the thing to watch.
One caveat should not be buried: Podqi sells brand protection, wrote its own prompts, and one tested brand — Madhappy — is a Podqi client. That does not invalidate the findings, but the study was designed by a company with a commercial interest in the problem it measured, using its own prompts rather than real user behaviour. Treat the headline numbers as a serious signal, not a precise measure of what ordinary shoppers experience.
Even with that caveat, the underlying mechanism is easy to believe. Chatbots retrieve from the live web, which is full of expired domains, lookalike stores and SEO spam. A system ranking sources by relevance and freshness, without checking whether a domain still belongs to the brand it claims to represent, will inevitably surface some of them — and the 351 fake and 367 dead domains here are evidence of that failure mode at scale.
The right response from the platforms is not to stop citing the web — that would defeat the purpose of shopping search — but to verify the domains they recommend: checking registration history, flagging recently re-registered brand domains, and warning users when a recommended site is not the brand's official store. Until that happens, shoppers should treat chatbot recommendations with the same caution as any unfamiliar link.
For SEOs, this study is a warning about the new trust economy in AI search. As chatbots become shopping assistants, the question of which sites they recommend becomes a commercial question, not just a visibility one. The brands that suffer when a chatbot recommends a fake store are the brands whose official sites are not being surfaced clearly enough — and part of the fix is the same work SEOs already do: making the official site unmistakably authoritative, well-linked and easy for an AI to identify as the real one.
For link builders, the lesson is that authority is now a safety signal. A brand with strong, genuine links and a clear digital footprint is easier for an AI to recognise as legitimate, and harder for a fake store to impersonate in the rankings that matter. The fake stores in this study were not outranking the real brands on authority — they were being surfaced because the AI's retrieval did not distinguish them. Building unmistakable authority for the real site is the best defence against that.
There is also a direct opportunity here. Brand-protection and domain-monitoring content — explaining how to spot fake stores, how to check whether a domain changed hands, how to protect a brand in AI search — is exactly the kind of genuinely useful, linkable content that earns citations. The Podqi study itself is proof that this topic attracts coverage; the next good study, or the best practical guide, will too.
Sources: The Verge, November 12, 2026.
Related reading on this site: OpenAI's Aura Safety Check Can Block Your Site From ChatGPT Browsing, Wikipedia Leads ChatGPT Citations at 10.1% in Ahrefs' November Ranking, Direct Answers Win AI Citations: 193,507 Answers Studied, How Often Sites Block AI Crawlers: 14,318 Sites Measured, AI Overviews Now Cover Over 80% of UK Keywords, SISTRIX Finds, and The Traffic Pool Is Shrinking — and That Changes Everything.
Linkbaits helps you create content that earns citations and backlinks. Start with the toolkit and build your first linkable asset this week.
Start free trial