Understanding the Risk of AI Chatbots in Online Shopping
Popular artificial intelligence chatbots including ChatGPT, Google's Gemini, and Perplexity frequently surface potentially risky websites when users ask where to buy products online. A comprehensive study conducted by intellectual property protection firm Podqi examined the safety of recommendations provided by these conversational tools. The investigation found that approximately 9% of the 6,795 answers analyzed included a risky website. These hazardous destinations ranged from fake storefronts and suspected counterfeit sellers to inactive or re-registered brand domains that could exploit unsuspecting consumers.
As consumers increasingly rely on conversational search engines for daily recommendations, the presence of deceptive links introduces significant cybersecurity and financial threats. The findings carry heightened urgency as they arrive just ahead of the holiday shopping season. During peak shopping periods, consumers searching for lower prices or hard-to-find items often turn to unfamiliar retailers without exercising their usual caution. Understanding how these tools select and display retail sources helps users protect themselves against online fraud and identity theft.
How Different Chatbots Compared in Safety Tests
The Podqi study evaluated three major AI platforms and uncovered distinct safety records among them. Perplexity surfaced risky sites in 15% of its answers, making it the most frequent provider of dangerous links among the tested services. ChatGPT followed with risky sites appearing in 8% of its answers, while Google's Gemini recorded the lowest rate at 4%. Researchers counted a site whether the chatbot recommended it, mentioned it without comment, or actively warned shoppers against using it.
Chatbots handled the identified hazards differently when presenting them to users. ChatGPT offered warnings most often, flagging the risk in 24% of its answers that included a risky site. Across all three services combined, 57% of answers that surfaced a risky site recommended it as a viable place to buy. Another 34% mentioned the domain without any comment, and only 9% warned shoppers away from the dangerous link. Podqi reviewed 4,437 unique domains during the research process and classified 718 as risky, which included 351 fake sites and 367 inactive ones.
Why Regional Store Questions Carried the Highest Risk
The type of prompt used heavily influenced the likelihood of encountering a dangerous recommendation. Questions concerning international shipping and regional storefronts surfaced risky sites in 16% of answers, representing the highest rate among the five distinct question categories tested. Chatbots sometimes invented or listed fake regional sites for brands that did not operate local storefronts in those specific regions. For example, in one test, Perplexity identified the wrong website as Stussy's official store and explicitly recommended it as a source of authentic products.
Brand size also played a role in the frequency of erroneous or malicious links. Questions about larger, globally recognized brands surfaced risky sites about half as often as questions focused on smaller brands. Podqi crafted the prompts and used data from artificial intelligence search tracking company Gauge to collect answers from the logged-out versions of the three services. The testing procedure ran in two rounds during September and covered 151 distinct brands, including major names like Lululemon, Birkenstock, Stussy, Le Creuset, and Madhappy, which is a Podqi client. The published figures directly reflect the answers generated by those specific test prompts.
Responses From Technology Companies
Representatives from the affected technology firms addressed the findings following the publication of the study. Google spokesperson Jessica Levis Boord told The Verge that Gemini can make mistakes and noted that Google displays source links so users can independently check its answers. OpenAI spokesperson Taya Christianson stated that the company is actively improving its safeguards to reduce the chances of deceptive websites appearing in shopping results. Perplexity spokesperson Jesse Dwyer pointed to the company's performance on other research benchmarks and mentioned that the service does not claim perfect accuracy.
These corporate statements highlight the ongoing technical challenges companies face in filtering out fraudulent web domains. Large language models generate text based on probabilistic patterns rather than verified fact-checking databases, making them vulnerable to SEO manipulation by malicious actors. As bad actors create sophisticated replica domains that mimic legitimate brands, AI filters struggle to distinguish genuine retailers from clever counterfeits. Industry observers note that closing these security gaps requires continuous updates to safety classifiers and stricter domain validation protocols.
Practical Safety Takeaways for Online Shoppers
Shoppers navigating online retail spaces must adopt rigorous verification habits when using conversational search tools. Podqi co-founder and CEO Trevor West advised consumers to thoroughly verify a website before entering personal information or sensitive payment details. When searching for regional stores or international shipping options, Podqi recommends starting directly from the brand's main website rather than trusting third-party conversational links. Checking uniform resource locators carefully and looking for secure connection indicators remain essential steps for safe e-commerce engagement.
Consumers should also remain skeptical of deep discounts or unusual domain endings associated with well-known brands. Cross-referencing chatbot suggestions with established review platforms and official corporate directories minimizes the risk of falling victim to phishing schemes. By maintaining vigilance and treating conversational recommendations as starting points rather than guaranteed safe transactions, shoppers can better protect their personal data and financial assets during the busy retail season.