A test of 6,795 shopping answers found that Gemini, ChatGPT and Perplexity all surfaced websites that were fake, inactive, re-registered after a brand abandoned them or selling knockoffs. Brand-protection company Podqi ran the study across 151 retailers during two periods in September.
Gemini mentioned or linked to a risky site in 4% of tested answers, ChatGPT in 8% and Perplexity in 15%. ChatGPT warned about 24% of the risky sites it surfaced, the highest warning rate among the three. Smaller brands appeared more exposed, while questions about regional stores produced risky results more often than other prompt categories.
The figures need careful interpretation. Podqi created the prompts rather than sampling real conversations, and one tested brand is its client. A mention also counted when a chatbot warned that the site was risky. The study therefore identifies a failure mode but does not measure the share of everyday shopping sessions that lead to fraud.
The practical risk grows when a plausible citation is treated as verification before entering payment details. Google and OpenAI said they continue to improve information quality and deceptive-site safeguards. Perplexity emphasized its broader research benchmarks while acknowledging that no system is perfectly accurate.