The new front line of e-commerce fraud: how AI chatbots are sending shoppers to fake websites

If you asked an artificial intelligence assistant where to buy a specific product in recent weeks, there was a significant chance it directed you to a site that doesn't exist — or operates as a scam storefront. A study conducted by brand-protection firm Podqi and shared with The Verge revealed that the three leading chatbots tested recommended dangerous retail websites in up to 15% of cases, a figure that raises urgent questions about the reliability of these tools as the holiday shopping season approaches.

Podqi's research analyzed 6,795 responses generated by ChatGPT, Gemini, and Perplexity to questions about purchases from 151 brands, including Lululemon, Birkenstock, Stüssy, Madhappy, and Le Creuset. The prompts were designed by the company itself, covering questions such as "what are the best sites selling this brand," "where can I find it online," and "are there regional or international stores." The study defined a website as "risky" if it appeared to impersonate a legitimate retailer, sell counterfeits, or was an expired brand domain that had been re-registered by third parties.

Results by platform

The results varied significantly across the three tools. Perplexity had the worst score: in 15% of its responses, it presented risky sites — about one in seven. ChatGPT came second with 8%, and Gemini had the lowest rate at just 4%. However, even the lowest figure represents thousands of potentially deceived visitors.

An interesting aspect of the study was the variation in chatbot behavior when a risky site was identified. ChatGPT was the most proactive: in 24% of cases where it presented a risky site, it warned the user about the risk. This alert rate is the highest among the three tools. Gemini and Perplexity, meanwhile, tended to list the site without any warning.

The study also revealed that vulnerability varied according to brand size. Smaller brands were affected approximately twice as often as larger ones, possibly because they have a less robust online presence — leaving an information vacuum that fake sites can easily exploit.

The regional problem

Questions about regional sites consistently produced the worst results. In several cases, chatbots listed regional sites for brands that have no official regional presence. One documented example shows Perplexity identifying the wrong page as Stüssy's official website and describing it as "the best starting point for guaranteed authentic items" — a particularly serious error, as it confers authority on a fraudulent site.

This pattern makes sense from a technical perspective. AI models are trained on web data that includes not only official sites, but also clones created by scammers. When a model "decides" which URL to present, it is essentially voting on relevance without verifying the domain's legitimacy.

The broader context

This study fits into a wider trend of AI-driven online fraud. Mastercard warned the previous week that generative AI is making it easier for criminals to create convincing fake online stores at scale. The scam-checking service Ask Silver had documented earlier cases where ChatGPT directed users to clone sites of Russell & Bromley and Dunelm, with discounts of up to 80% — a classic red flag of e-commerce scams.

The UK's National Trading Standards, through its head Louise Baxter, warned that "consumers are increasingly turning to AI tools for advice and recommendations, but criminals are adapting just as quickly." The observation reflects a structural problem: as AI chatbots become the first product discovery interface for consumers — replacing traditional search engines — they inherit and amplify the risks of the entire web.

What to expect

Podqi's study was published with three weeks' lead-up to the holiday shopping season, when AI use for product discovery is expected to spike. Data collected in September showed a risk reduction between the first and second testing waves (from 12% to 7%), suggesting providers are slowly correcting the problem — but that progress remains insufficient to protect millions of consumers.

The fundamental question is whether the AI industry can improve URL validation in time to make the holiday shopping season safer for everyone — or whether consumers will have to learn, once again, that the convenience of an AI assistant is not synonymous with absolute reliability.

Sources: The Verge, Retail Risk, Podqi Research

✓ Independent sources cross-checked and verified before publishing