Skip to content

    AI Chatbots Recommend Risky Fake Retail Sites in Up to 15% of Answers

    AL
    Aria Lin

    November 13, 2026

    Perplexity named the wrong page as Stüssy's official website and called it "the best starting point for guaranteed authentic items," according to a study by the IP protection firm Podqi that analyzed 6,795 chatbot answers across 151 brands. The Stüssy answer is the vivid case,

    AI Chatbots Recommend Risky Fake Retail Sites in Up to 15% of Answers

    Perplexity named the wrong page as Stüssy's official website and called it "the best starting point for guaranteed authentic items," according to a study by the IP protection firm Podqi that analyzed 6,795 chatbot answers across 151 brands. The Stüssy answer is the vivid case, but the buried lead is in the fine print: Podqi's percentages count any answer that merely mentions a risky site, including answers that warned the user away from it, so the figures measure exposure rather than proven harm. The practical question is how often a chatbot's confident link points somewhere a card number should never go.

    What Happened

    Podqi, which sells brand and intellectual property protection, ran the study and shared it exclusively with The Verge. It tested ChatGPT (OpenAI), Gemini (Google) and Perplexity across 151 brands, a mix of small and large companies that included Lululemon, Birkenstock, Stüssy, Madhappy and Le Creuset. Testing ran in two periods, both in September 2026, and produced the 6,795 answers that Podqi then analyzed.

    The study used three question templates, which asked chatbots for the top places to buy a brand, where to find it online, and whether it ships internationally and has regional websites.

    As described in the reporting, Podqi defined a "risky" site in three ways: fake websites impersonating real retailers or selling knockoffs, old brand domains re-registered by someone else, and dead sites that are inactive but could be purchased and reused. A "mention" counted whether the chatbot recommended the site, displayed it without comment, or warned the user about it.

    One inconsistency runs through the coverage. The headline and deck refer to "conversations," while the body refers to "answers." Those are different units, and the difference matters when a reader sees "as many as 15 percent" in a headline.

    Why It Matters

    A shopper who trusts a chatbot's link may enter personal or credit card details on a fake site. By Podqi's count, risky sites appeared in 15% of Perplexity's answers, 8% of ChatGPT's and 4% of Gemini's. Perplexity's rate is nearly four times Gemini's, a gap wide enough to matter for anyone choosing which assistant to ask where to shop.

    Size matters too. Queries about larger companies surfaced risky sites about half as often as queries about smaller brands, so smaller brands carry more of the risk. The reporting also indicates that the regional-site question was the most problematic of the three, and that chatbots sometimes listed fake regional sites for brands that have none.

    A smartphone with a dark screen lies beside a chip credit card on a wooden table while a finger reaches toward the card.

    That pattern points to a possible structural cause rather than random error: when a model is asked for something that may not exist, such as a regional storefront, it can supply a plausible-looking domain anyway. Researchers call this hallucination (a language model generating plausible but false output), and the survey by Huang et al. cited below catalogs it as a known weakness of large language models. The Podqi study does not establish that hallucination, rather than retrieval of real but malicious pages, produced the specific results.

    Trevor West, Podqi's CEO and cofounder, told The Verge: "Part of this piece, too, is making sure that the consumers know, even though you are being fed a result in a particular chatbot, it is always best to verify the source of that before entering personal information or credit card information."

    Competitive Landscape

    The three chatbots ranked in a clear order on how often risky sites appeared in answers:

      • Gemini (Google), best: risky sites in 4% of answers.
      • ChatGPT (OpenAI), middle: risky sites in 8% of answers, and the best at warnings, cautioning users about each risky site it surfaced 24% of the time.
      • Perplexity, worst: risky sites in 15% of answers.

    The companies responded by email. Jessica Levis Boord, a Google spokesperson, said: "We take information quality seriously, and are happy to see that Gemini performed best among the products tested... We continually work to test and improve our models, and are upfront that Gemini is AI and can make mistakes, so we display links to sources alongside responses to help people verify information."

    OpenAI spokesperson Taya Christianson said: "Our goal is for ChatGPT to connect people with reliable places to shop... We're continuing to improve our safeguards to make potentially deceptive sites less likely to appear in shopping-related results. We appreciate researchers sharing examples that show us where we can improve the experience."

    Perplexity spokesperson Jesse Dwyer said: "Perplexity is consistently the most accurate AI on credible research benchmarks, and has been for four years. We never claim to be 100% accurate, but we do claim to be the only company relentlessly focused on it." The statement does not address the 15% figure.

    The Bigger Picture

    A hand holds a cheap white feature phone beside a torn-open cardboard box and a crumpled paper slip on a doorstep mat.

    The study's credibility has limits. Podqi is an IP protection firm with a commercial interest in showing that the problem exists, and Madhappy, one of the brands in the sample, is a Podqi client.

    The blended count is the largest interpretive gap. Because a warning counts as a mention, ChatGPT's 8% includes answers that told users to be careful, and the study does not separate recommendations from displays or warnings. It also measures exposure, not harm: nothing in the reporting shows that any user lost money.

    The broader trend is real regardless. The Verge notes that AI companies market chatbots for researching, comparing and discovering products and retailers, which gives counterfeit and phishing sites a new route to buyers.

    What's Next

    For readers, the steps are simple: check the domain before paying, open the source links a chatbot displays rather than trusting the summary, and go to the brand's own site directly. Take extra care with regional storefronts.

    Three things are worth watching. OpenAI says it is improving safeguards; Perplexity's statement does not mention changes. Independent replication would show whether the percentages hold.

    For a product manager or developer building shopping features on top of a chatbot, the 4%-to-15% spread is a reason to ask vendors what domain verification sits between the model's answer and the link a customer clicks, and to treat displayed source links as a required part of the interface rather than an optional extra.

    The firm that sells brand protection has found a problem in the tools that are becoming the front door to retail, and its own incentives are part of the story. Until Podqi's data can be independently checked, the numbers are best read as a credible warning, not a measurement. A chatbot that says "guaranteed authentic" is still only guessing at a domain.

    -- Aria Lin, Enterprise Technology Analyst


    Sources: Mit · Mit · The Verge · arXiv

    More on Revuzia