The dangerous part of an AI shopping answer may arrive after the useful part. A chatbot can help narrow a product choice and then attach a store address that the shopper has never checked. At that point, confidence in the recommendation can turn into confidence in a merchant, even though those are separate judgments.
A Podqi audit of AI shopping answers makes that handoff worth examining. The brand-protection company reports risky websites appearing in about 9% of its tested answers. The evidence supports a specific concern about referral quality; it does not show that one in eleven AI-assisted purchases ends in fraud.
An answer rate is not a loss rate
Podqi analysed 6,795 answers about 151 brands across ChatGPT, Gemini and Perplexity, using two waves of September 2026 fieldwork. Its definition of a risky destination includes suspected brand impersonation or counterfeits and inactive sites where a shopper could buy nothing. Those outcomes have different consequences.
The widely reported 15% figure applies to Perplexity's answers in this test, rather than all assistants or all shopping activity. These are observations from a defined prompt set and period. They are not a current universal ranking, a consumer's personal probability of losing money or evidence that each flagged domain completed a fraudulent transaction.
Podqi also sells brand-protection services, a commercial interest readers should retain when evaluating its framing. That does not invalidate the findings, but it strengthens the case for examining the method and independent evidence instead of treating the headline percentage as a complete measure of risk.
Netcraft's separate research, published in August, provides such corroboration of the broader mechanism. Its June 2026 test found malicious links in 1.7% of 2,905 responses to queries about brand login pages. Different questions and a narrower malicious-link outcome mean the two studies cannot be combined into a single rate or read as a trend.
The handoff to a merchant is where authority changes
Product discovery and seller authentication answer different questions. A recommendation may identify an appropriate item while getting the seller's regional website wrong. The financial exposure begins when the user acts on that destination, supplies credentials or enters payment information.
A polished explanation can obscure this boundary. The assistant's surrounding prose may be coherent and useful even when a cited domain belongs to an unrelated operator. A link labelled as official is still a claim about ownership. Its presence in an answer does not establish that the brand controls the destination.
The UK National Cyber Security Centre's shopping guidance explains how criminals copy legitimate website designs and use deceptive domains. It recommends checking unfamiliar shops and reaching known organisations through their official address. For an AI-supplied regional link, an independently reached brand site can supply a better starting point than the assistant's description alone.
Connection security is another separate test. The US Federal Trade Commission notes that HTTPS means the connection is encrypted, not that the seller is legitimate. A secure connection to an impersonator still sends information to the wrong party. A familiar visual design and a browser security indicator cannot substitute for identifying the merchant.
Payment protections belong to the transaction you actually make
The FTC advises researching the seller and keeping order records. Its US guidance explains that credit-card billing errors can be disputed and warns against sellers that insist on hard-to-recover payment methods. Those protections should not be translated into a worldwide promise that every disputed purchase will be reimbursed.
The NCSC similarly advises checking the terms of payment services. The available remedy depends on the payment route, applicable rules and circumstances. A buyer who follows an AI link has not acquired an additional guarantee from the assistant merely because it recommended the store.
This creates a practical sequence: establish who is receiving the money, understand the transaction terms, and preserve the evidence needed if something goes wrong. The checks address different failure modes. Verifying identity reduces misdirection; reading delivery and return terms clarifies the agreement; records support a later complaint.
The benefit of an assistant can remain at the earlier stage: comparing product features or helping formulate questions. Moving from that help to payment requires a fresh trust decision. Convenience is useful only when it does not silently remove the checks that protect the purchase.
Retailers inherit a distribution problem as well as a security problem
For a legitimate retailer, a wrong destination can mean a lost sale, a confused support request or a reputational complaint about goods it never sold. These are plausible commercial channels, not losses quantified by the cited audits. The studies do not justify multiplying an answer rate by total online spending.
Clear official-domain and authorised-seller information could make referrals easier to verify. Monitoring the destinations associated with a brand also tests whether visibility in an answer leads to the intended store. Appearing prominently is less valuable if the next click sends the customer elsewhere.
Repeated audits with disclosed methods, stable comparisons and better destination checks would show whether the problem is improving. Documented consumer outcomes would be needed to estimate financial harm. Until then, the strongest conclusion is narrower and actionable: an AI-generated shopping link is a referral to investigate, and the merchant receiving the payment remains a separate counterparty.