Perplexity showed shoppers risky store sites in 15% of answers, ChatGPT in 8%

Ask an AI chatbot where to buy a brand online and the answer can include a fake store, a knockoff seller or an old brand domain that now belongs to someone else. A new study that put ChatGPT, Gemini and Perplexity through thousands of shopping questions found all three surfaced such sites, and smaller labels were the easiest targets. If a chatbot is doing your gift research this season, treat the link it hands you with the caution you would give a stranger’s text message.

The research comes from Podqi, a company that protects brands’ intellectual property, and was shared exclusively with The Verge. Podqi asked the chatbots questions any shopper might type: which websites are best for buying a given brand, where to find it online, and whether it ships abroad through regional sites. Its definition of a risky site was broad by design, covering pages that impersonate real retailers, shops selling counterfeits, expired brand domains re-registered by strangers, and dead sites that anyone could buy and reuse.

Perplexity did worst, showing a risky site in 15 percent of its answers. ChatGPT came in at 8 percent and Gemini at 4 percent, across 6,795 answers about 151 brands that ranged from Lululemon, Birkenstock and Le Creuset to streetwear labels such as Stüssy and Madhappy. Those figures count every appearance, including cases where a chatbot named a risky site in order to warn against it. ChatGPT was the most cautious of the three, yet it attached a warning to only 24 percent of the risky sites it surfaced.

Questions about regional storefronts caused the most trouble. Asked for country-specific shops, the chatbots sometimes listed regional sites for brands that have none. In one exchange, Perplexity named the wrong page as Stüssy’s official website and called it “the best starting point for guaranteed authentic items.” Queries about large companies returned risky sites about half as often as queries about smaller ones, so the less famous the brand, the easier it is for an impostor to slip into the answer.

The format makes the problem worse. A search results page lists several options and leaves the choice to you, while a chatbot names one or two shops in a confident sentence that reads like a personal recommendation. AI companies actively pitch their assistants for product research and price comparison, which is exactly the moment a shopper is about to type in a card number.

The study has clear limits. Podqi wrote its own prompts instead of analysing real user conversations, and one of the tested brands, Madhappy, is a Podqi client. The company also sells brand protection, so it benefits when impersonation looks like a serious threat. The results do not show how many people clicked a risky link or lost money, and they capture models that change every few weeks.

The companies answered on the record. Google said it was pleased Gemini performed best and noted that it displays source links so people can check its answers. OpenAI said it is improving safeguards to make deceptive sites less likely to appear in shopping results. Perplexity defended its accuracy record and said it has never claimed to be 100 percent accurate.

The finding fits a wider pattern. In September the cybersecurity company Eset warned that scammers publish floods of keyword-stuffed pages designed to be scraped by AI tools and repeated back to users as trusted advice. Podqi co-founder Trevor West’s advice is simple: “it is always best to verify the source” before entering personal or card details. The tests ran in two rounds during September, and the timing matters for shoppers everywhere these chatbots operate, with Black Friday on November 27 and Cyber Monday on November 30 sending millions of first-time visitors to unfamiliar stores.

Tags: OpenAI, chatgpt, Gemini, AI chatbots, online shopping, Perplexity