Consumers using artificial intelligence chatbots to research products and find online stores are frequently being directed toward imposter storefronts and other hazardous web destinations. A comprehensive new study conducted by intellectual property protection firm Podqi has revealed that major AI conversational engines routinely cite and display questionable URLs in response to everyday shopping queries. The research underscores emerging consumer risks as digital commerce increasingly integrates conversational search tools ahead of major shopping periods.
Key takeaways
- Podqi evaluated ChatGPT, Gemini, and Perplexity across 6,795 answers concerning 151 large and small brands.
- Perplexity registered the highest rate of problematic recommendations, displaying risky sites in 15 percent of its responses.
- Google's Gemini performed best in the evaluation, featuring risky URLs in only 4 percent of its answers.
- Queries involving smaller companies were found to trigger fake or unverified retail sites about twice as often.
Evaluating chatbot accuracy across retail queries
The study was conducted over two separate testing periods in September, using a specialized tracking tool to monitor AI answers across multiple platforms. Podqi formulated specific shopping prompts covering 151 companies, ranging from major global names like Lululemon, Birkenstock, Stüssy, and Le Creuset to smaller businesses like Madhappy. The tested prompts included direct consumer inquiries such as where to find specific brands online, which websites are best for purchasing them, and whether brands offered international shipping options.
Over the course of 6,795 analyzed answers, every single AI engine evaluated cited risky sites. These hazardous destinations included fake websites impersonating legitimate retailers, counterfeit operations selling knockoffs, dead domains that could be purchased and reused by malicious actors, and old brand domains that had been reregistered by third parties. Perplexity performed the worst among the tested systems, displaying risky sites in 15 percent of its answers. ChatGPT landed in the middle, showing questionable links in 8 percent of responses while providing warnings about those sites 24 percent of the time. Google's Gemini recorded the lowest rate, presenting risky URLs in 4 percent of answers, which included mentions, recommendations, or explicit warnings.
Risks to consumers and platform accountability
The inclusion of fraudulent domains in AI-generated shopping recommendations introduces significant financial and security risks for unsuspecting users. Particular prompt categories were more prone to triggering hazardous results, such as inquiries regarding regional brand websites. In some instances, the AI engines even fabricated fake regional pages for brands that do not maintain any such presence. For example, Perplexity misidentified an incorrect page as the official site for Stüssy, erroneously labeling it as a guaranteed source for authentic items.
Technology companies have actively marketed their conversational platforms as convenient assistants for product discovery, price comparison, and retailer research. However, experts warn that consumers must exercise caution. Podqi CEO and cofounder Trevor West emphasized that shoppers should always independently verify sources before submitting personal data or credit card details, regardless of how confident a chatbot's response appears.
Representatives from the affected tech firms responded to the findings by outlining ongoing efforts to enhance model safety. Google spokesperson Jessica Levis Boord stated that the company takes information quality seriously and noted that Gemini displays source links alongside responses to assist users with verification. OpenAI spokesperson Taya Christianson explained that developers are continuously refining safeguards to reduce the likelihood of deceptive sites appearing in shopping results. Perplexity spokesperson Jesse Dwyer highlighted that his company remains relentlessly focused on research accuracy while acknowledging that absolute perfection is difficult to guarantee.
Comments