AI Chatbots Can Send Shoppers to Fake Stores, Study Finds

A study by intellectual property firm Podqi found that popular AI chatbots like ChatGPT, Google's Gemini, and Perplexity can recommend potentially dangerous websites when asked for product purchase locations. Approximately 9% of 6,795 responses contained risky sites, which ranged from counterfeit stores to inactive brand domains.

Perplexity generated risky site recommendations in 15% of its answers, ChatGPT at 8%, and Gemini at 4%. ChatGPT was most likely to warn users about the risks, appearing in 24% of its risky site answers. In 57% of cases, the chatbots suggested using the risky sites. Nearly 60% of the dangerous sites recommended by the chatbots were categorized as risky by Podqi.

The highest-risk queries were those about international shipping and regional storefronts, which produced risky sites in 16% of answers. Sometimes, the chatbots listed nonexistent regional sites for brands that had no such stores. One instance involved Perplexity mistakenly recommending an incorrect website as the official store for a brand.

Larger, well-known brands were queried about risky sites less frequently than smaller ones. Podqi conducted the tests using prompts from Gauge data and in two rounds in September, covering 151 brands, including Lululemon, Birkenstock, Stüssy, Le Creuset, and Madhappy. Google's Gemini can make errors, according to a Google spokesperson, and Google shows source links to verify the chatbot's answers.

OpenAI is working on improving safeguards to reduce deceptive websites in shopping results. Perplexity acknowledged its performance on other research benchmarks and stated it does not claim perfect accuracy. As the holiday shopping season approaches, Podqi advises shoppers to check websites for personal information or payment details, especially when dealing with regional stores.