Study: AI Chatbots Recommend Fake Websites to Users
A Podqi study found that ChatGPT, Gemini, and Perplexity displayed risky websites in 9% of responses to shopping-related queries. Perplexity did this most frequently.
The authors tested responses from ChatGPT, Gemini, and Perplexity to five shopping-related questions for 151 brands, including queries about official and regional websites. The testing was conducted in two waves in September 2026: a total of 6,795 responses and 4,437 domains were analyzed.
Websites were considered risky if they impersonated a brand, could sell counterfeit products, or were inactive. Such addresses appeared in 15% of Perplexity’s responses, 8% of ChatGPT’s responses, and 4% of Gemini’s responses. On average, chatbots recommended a risky website as a place to shop in 57% of such responses, mentioned them without comment in 34% of cases, and warned about them in only 9%.
The most problematic queries were about brands’ regional websites: a risky address appeared in 16% of responses. The systems sometimes provided websites for countries where the brand had no regional store at all. According to the study, out of 718 risky domains, 351 were linked to counterfeits, while 367 were inactive. The authors also noted that in the second wave, the share of such responses decreased from 12% to 7%.
