Asking a chatbot where to buy a brand online feels safer than clicking a search advert, but a new study suggests the answers are not always clean. Podqi, a firm that protects brands' intellectual property, found that ChatGPT, Gemini and Perplexity all pointed shoppers to risky websites, in up to 15% of answers. With the holiday season close and more people shopping through AI, that is a practical warning rather than an abstract one.
The study, shared exclusively with The Verge, asked shopping questions about 151 brands, from Lululemon, Birkenstock and Stüssy to Le Creuset and Madhappy, which is a Podqi client. Prompts included where to buy a brand online and which regional websites it runs. Across 6,795 answers collected over two test periods in September, a risky site meant one of three things: a fake shop impersonating a real retailer or selling knock-offs, an old brand domain re-registered by someone else, or a dead site that anyone could buy and reuse.
The results separate the three. Gemini showed or mentioned a risky site in 4% of answers, ChatGPT in 8% and Perplexity in 15%. Mentions include recommending a site, listing it without comment and warning about it; ChatGPT was best at adding a warning, doing so 24% of the time. Questions about regional websites were the most likely to produce bad links, and the chatbots sometimes invented regional sites for brands that have none. In one case Perplexity named the wrong page as Stüssy's official site and called it the best place for guaranteed authentic items. Smaller brands were hit harder: questions about larger companies surfaced risky sites about half as often.
For readers, the advice is simple and comes from Podqi's chief executive, Trevor West: check the address of any shop a chatbot gives you before entering card details. It also matters for agents like OpenAI's Dots, which are starting to buy things directly. An agent that trusts the wrong domain could make the mistake for you, which is why per-purchase approval matters for now.
The study has clear limits. Podqi wrote its own prompts rather than using real conversations, it has a commercial interest in brand protection, and one of the brands tested is its client. Google, OpenAI and Perplexity responded to The Verge; Perplexity said it never claims to be 100% accurate. The results are a snapshot from September, and chatbots' sources change often.