What "risky" actually covers
The headline rate bundles different things. A flagged site could be an apparent impersonator or a suspected counterfeit seller. It could be an old brand domain that lapsed and was re-registered by a third party. Or it could be a dead domain: inactive, a destination where a shopper can buy nothing. The split matters: of the 718 flagged domains, Podqi classified 351 as fake and 367 as dead. A dead domain in an answer is a stale citation today. Podqi's own argument for counting it is that anyone can re-register a lapsed domain, so an address a chatbot keeps recommending could become a counterfeit storefront later. Counting both kinds in one percentage makes the problem sound uniformly live; among the flagged domains themselves, slightly more than half were dead at test time.
The counting rule widens the number further. Podqi counted an answer as exposing a shopper if a risky site was shown or mentioned, including answers where the chatbot warned about the site. A warning and a recommendation land in the same bucket. The more useful figure sits underneath: when a risky site appeared, the chatbot recommended it as a place to shop 57% of the time, said nothing 34% of the time, and warned 9% of the time. Recommended roughly 6.3 times as often as warned against is the finding worth keeping. The 15/8/4 figures are comparative exposure rates under Podqi's prompts and risk definition, not general accuracy rankings for the three services.
Where the answers went wrong most
Questions about regional stores and international shipping produced risky sites in 16% of responses, the riskiest category in the test. In some cases the chatbots supplied regional sites for brands that do not operate one, which is a different failure from citing a bad link: the answer filled a gap with a plausible address. Smaller brands were exposed at roughly twice the rate of large ones, as reported by The Verge; the study does not establish why.
The Stüssy case, reported by The Verge, is the clearest documented incorrect recommendation. Perplexity named an incorrect page as the brand's official site and described it as a source of authentic products. That is not a dead link. It is a confident referral to the wrong store, delivered weeks before the heaviest shopping period of the year.
The limits, stated plainly
Podqi sells brand-protection services. A study showing that brand impersonation is common and under-policed is a study that describes demand for its own product, and one of the 151 brands tested, Madhappy, is a Podqi client, as reported by The Verge. Neither fact invalidates the domain counts, which are checkable in principle, but both belong next to the headline.
The prompts were Podqi's own, five template questions run through an answer-tracking tool (data via Gauge), not a sample of real shopper conversations. The study measured exposure: it did not track whether anyone clicked, bought, or lost money. And the snapshot is already aging. The overall risky rate fell from 12% in the first September round to 7% in the second, inside the same study. The observed rate changed between rounds; the study does not isolate the cause, so neither figure is a fixed defect rate for any of the three services.
Company responses, as reported by The Verge: OpenAI did not dispute the findings and said it is continuing to improve safeguards for shopping results; Google pointed to its best-in-test result; Perplexity's spokesperson Jesse Dwyer said the company never claims to be 100% accurate and pointed to its focus on accuracy. The study measured shopping citations; the replies speak to accuracy in general.
What a shopper should actually do with this
Treat a chatbot's store recommendation as a lead, not a destination. Reach the store through the brand's official site, and verify that the retailer is linked or identified there before paying. Treat regional-store claims with the most suspicion, because that is where the test found answers naming regional sites for brands that do not operate one. The study gives no reason to stop asking chatbots where to buy things. It gives a specific reason not to pay at a store a chatbot named without checking it against the brand's own site first.
ALFW has not re-run Podqi's test. All figures are Podqi's as published on its study page; company responses and the Stüssy case are as reported by The Verge.
Sources
- [1] Podqi, "AI search is recommending fake retail websites" (study, September 2026 fieldwork, published November 12, 2026)Read source
- [2] The Verge, "AI chatbots are recommending fake retail websites, study finds" (November 12, 2026) — company responses, the Stüssy case and the Madhappy client relationship