AI chatbots: when purchase recommendations point to dubious sites
A study by Podqi, reported by The Verge, shows that ChatGPT, Gemini, and Perplexity sometimes direct users to fake merchant sites. The risk mainly concerns small brands and queries related to regional stores.
ChatGPT, Gemini, and Perplexity increasingly serve as gateways to the web, including for shopping. This is precisely where the problem arises: according to a study by Podqi reported by The Verge, these chatbots sometimes recommend risky sites capable of imitating real merchants, reactivating abandoned domains, or selling counterfeits. The issue matters because a response presented as reliable can lead a user to enter their banking details in the wrong place.
Quand l’IA confond boutique légitime et piège en ligne
The test covered about 151 brands, from small shops to better-known names like Lululemon and Birkenstock. Researchers asked questions such as "What are the top websites for buying [brand]?", "Does [brand] ship internationally?", or "What are their regional websites?". Result: Perplexity showed dubious sites in 15% of its answers, ChatGPT in 8%, and Gemini in 4%.
The finding does not stop there. Among responses containing a risky site, 57% recommended it as a place to buy, 34% mentioned it without warning, and only 9% explicitly warned the consumer. ChatGPT did add a warning in 24% of its responses, which remains the highest score of the trio, but this is not enough to dispel the risk. Small brands also seem more exposed, especially on queries about regional sites.
Des réponses sûres en apparence, mais pas toujours vérifiées
Trevor West, CEO of Podqi, sums up the issue: "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," he told The Verge. The recommendation is simple, almost banal. Yet it remains necessary, as these tools often give the impression of deciding without hesitation, while poorly explaining how they arrive at their results.
This opacity distinguishes them from a classic search engine, easier to cross-check even if it can also be manipulated. Podqi is not alone in sounding the alarm: the anti-scam service Ask Silver also found that ChatGPT recommended cloned versions of closed sites designed to steal banking information. Another embarrassing demonstration for the industry: a BBC journalist showed he was able to convince ChatGPT to repeat a false identity of a hot-dog champion he invented. Generative models know how to respond quickly. They also know how to be fooled, sometimes with almost embarrassing ease.
The problem therefore goes beyond the simple quality of a commercial response. The more internet users use chatbots as a search interface, the more source selection errors can have a direct effect on purchase, payment security, and trust in these tools. In an already saturated web of fake content, AI models are not immune to ambient contamination.
By the numbers
- 151 brands — sample tested by Podqi, according to The Verge.
- 15% — risky responses observed with Perplexity.
- 8% — risky responses observed with ChatGPT.
- 4% — risky responses observed with Gemini.
- 57% — responses that recommended the dubious site as a store.