AI Shopping Assistants Can Point to Fake Websites, Study Finds

Introduction

AI chatbots are becoming shopping gateways. A user can ask one to compare products, locate inventory, or identify the official place to buy from a brand. But a test conducted by intellectual-property protection firm Podqi, and reported exclusively by The Verge, suggests that these recommendations can also lead shoppers toward imitation, abandoned, or otherwise risky websites.

What the test found

Podqi tested ChatGPT, Google Gemini, and Perplexity during two testing periods in September. It created shopping prompts for 151 large and small brands, including fashion and homeware companies, and analyzed 6,795 answers in total. The prompts asked where to buy a brand, which websites were best for it, and whether the brand had international or regional sites.

Podqi classified a URL as risky when it impersonated a genuine retailer, appeared to sell counterfeit goods, used an old brand domain that had been registered by someone else, or pointed to a dead site that could later be reused. All three chatbots mentioned or displayed URLs that fit one of those categories.

• Perplexity surfaced risky sites in 15% of answers.
• ChatGPT did so in 8% of answers.
• Gemini had the lowest rate, at 4%.
• ChatGPT provided a warning in 24% of cases involving a risky site.
• Queries about smaller companies produced risky results roughly twice as often as queries about larger ones.

The study counted a URL when a chatbot recommended it, displayed it without context, or warned users about it. That means the figures do not necessarily show that every system explicitly endorsed every risky site. Still, the appearance of a link in an authoritative-looking answer can create a powerful signal of legitimacy, especially for shoppers who do not know the brand’s official domain.

Regional websites are a particular weak spot

Questions about regional brand websites appeared to trigger risky links more often than other prompt types. Some answers also invented or listed regional sites for brands that did not have them. In one example, Perplexity identified the wrong page as Stüssy’s official website and described it as a good starting point for authentic products.

The danger goes beyond an ordinary factual error. A shopper may use such a page to create an account, enter an address, or submit card details. Domain ownership and website status can also change over time: a once-legitimate address may expire, while a convincing imitation can remain online long enough to attract traffic from AI-generated recommendations.

Smaller brands may face additional exposure because they have fewer widely recognized official channels and less online information for systems to cross-check. That does not prove that every small retailer is less safe, but it does show that brand size can affect how reliably a chatbot identifies the right destination.

What consumers and platforms should do

For shoppers, an AI answer should be treated as a lead, not as proof that a website is official. Before buying, users should compare the domain with links from the brand’s known social accounts or established official pages. They should avoid entering passwords, identity details, or payment information until the site’s ownership and security are independently confirmed.

The test also points to a platform responsibility. Shopping answers need stronger domain-status checks, source ranking, and handling of regional sites, authorized retailers, and lesser-known brands. Google said Gemini displays source links and that its systems are continuously tested, while OpenAI said it is improving safeguards against deceptive shopping sites. Perplexity emphasized its focus on accuracy without claiming perfection. The broader lesson is straightforward: a citation can show where an answer came from, but it cannot by itself prove that the destination is current, official, or safe.

Source: The Verge AI