AI shopping assistants are sending shoppers to fake storefronts

A shopper asked Perplexity where to buy authentic Stüssy products. The chatbot pointed them to a website it described as the brand’s official store and vouched for the authenticity of its merchandise. It had identified the wrong site.

The exchange was part of a study by intellectual property protection firm Podqi, shared with The Verge. Across 6,795 shopping-related answers from Perplexity, ChatGPT and Google’s Gemini, all three assistants surfaced web addresses that Podqi classified as risky. Questions about where to shop in a particular country were especially likely to produce one.

What the researchers measured

Podqi wrote shopping prompts covering 151 brands, including Birkenstock, Le Creuset, Lululemon and Madhappy, a Podqi client. During two testing periods in September, the researchers asked the assistants where to buy each brand online, whether it shipped internationally and which regional websites it operated.

Perplexity mentioned or linked to a risky site in 15% of its answers, compared with 8% for ChatGPT and 4% for Gemini. The category included impersonation sites and sellers of counterfeit goods, as well as former brand domains under new ownership and inactive addresses that someone could acquire.

A mention counted even when the assistant warned the shopper to avoid the site. ChatGPT issued such warnings in 24% of the instances in which it surfaced a risky address, the highest rate among the three. The figures therefore measure how often those addresses appeared in answers to Podqi’s test questions. They do not establish how often real shoppers received a fraudulent recommendation or followed one.

Why regional shopping questions matter

When researchers asked for a brand’s regional websites, the assistants sometimes supplied local storefronts for brands that had none. That distinction matters to someone shopping internationally. A brand may deliver to a country through its main store, and a shopper asking about shipping may have no reason to expect a separate national website until the assistant suggests one.

Brand size also made a difference in the test. Risky addresses appeared about half as often in answers about larger companies as they did in answers about smaller ones.

Google told The Verge that Gemini includes source links so users can check its answers. OpenAI said it was improving safeguards against deceptive shopping results. Perplexity pointed to its performance on research accuracy benchmarks and acknowledged that its answers can contain errors.

What retail software teams can test

For teams building AI-assisted shopping into a website or app, the Stüssy example suggests a specific test. When the assistant calls a store official, what evidence supports that claim? A working link establishes that a page is reachable. Verifying who operates it requires a trusted source, such as the brand’s own store directory.

One practical test case is a brand that ships to a country through its global store and has no local storefront. Give the assistant those facts, ask where a customer in that country can buy the product, and check the destination it presents. A good answer should point the shopper to the verified store and explain the shipping option. That is the detail the customer needs before deciding where to enter payment information.

Source: The Verge, reporting on research by Podqi.

Control F5 Team
Blog Editor
OUR WORK
Case studies

We have helped 20+ companies in industries like Finance, Transportation, Health, Tourism, Events, Education, Sports.

READY TO DO THIS
Let’s build something together