ChatGPT, Gemini and Perplexity sent shoppers to fake stores in a 6,795-answer test

Podqi shared the study with The Verge. It asked the three assistants shopping questions about 151 fashion and homeware brands, including Lululemon, Birkenstock, Stüssy, Madhappy and Le Creuset, and analyzed 6,795 answers. Every assistant surfaced risky destinations at least some of the time: stores impersonating real retailers, knockoff sellers, old brand domains that had been reregistered by someone else, and dead domains that anyone could buy.

The gap between products was large. Perplexity included a risky site in 15 percent of answers, ChatGPT in 8 percent and Gemini in 4 percent. ChatGPT was the most likely to attach a warning when it did surface one, doing so 24 percent of the time. Smaller brands were hit roughly twice as often as big ones, and the assistants sometimes listed regional storefronts for brands that do not have any.

A trivia mistake is embarrassing. A shopping mistake can mean a card number handed to a scammer. The study’s clearest example is Perplexity naming the wrong page as Stüssy’s official site and describing it as a place to buy guaranteed authentic items. That is the risk in one answer: a confident, friendly reply that lends the assistant’s credibility to a bad link.

Timing makes this urgent. More people are expected to shop through AI assistants this holiday season, and the companies are actively building commerce into their products. OpenAI has been adding ads and shopping partnerships to ChatGPT, from visual ads next to image results to grocery shopping with Albertsons’ Safeway inside the chatbot (our coverage). The more an assistant acts like a storefront, the more its link quality becomes a consumer-protection issue.

The study has real limits. Podqi wrote its own prompts instead of studying real user conversations. It ran the tests in September, and models change quickly. Its count includes mentions that came with a warning. Madhappy is a Podqi client, and the firm sells brand-protection services, so it benefits from a worrying result. All of that argues for treating the exact percentages as indicative rather than definitive.

Even so, the method is clear and the spread between products is wide enough to take seriously. Google and OpenAI told The Verge they are working on the problem. Perplexity pointed to its accuracy record. Independent replication with real user queries would settle how big the problem is in practice.

Scammers follow traffic. They once gamed search ads and SEO, and they will do the same for chatbot answers as those become the first stop for shopping. Smaller brands, which already struggle to police impostors, will feel it first. The fix is not to give up on AI shopping help, which is useful for comparing products and finding sellers. It is verified merchant data, clear warnings, and treating any claim that a site is official as something the system must check rather than guess. Podqi’s CEO offers the consumer version of that advice: verify a site before entering personal or card details. Follow our OpenAI and Google coverage for how the shopping features evolve.