What happened
The Verge reports on a study conducted by IP protection firm Podqi and shared exclusively with the publication. Over two testing periods in September, Podqi asked ChatGPT, Gemini and Perplexity shopping-related questions about 151 brands, ranging from large companies to smaller ones. Prompts included questions about the best websites for buying a brand, where to find it online, and whether it ships internationally or has regional websites. Podqi wrote the prompts itself and used a tool that tracks AI answers across chatbots; the test did not use real user conversations.
Across 6,795 answers, Podqi found that all three systems cited or showed risky sites. The study defined “risky” as fake sites impersonating real retailers or selling knockoffs, old brand domains that had been re-registered by someone else, and dead sites that were no longer active but could be purchased and reused. The rates were 15 percent of Perplexity answers, 8 percent of ChatGPT answers and 4 percent of Gemini answers. The count included a recommendation, an unannotated display, or a warning about a risky site, so it does not measure only cases where a chatbot explicitly endorsed one.
Podqi found that queries about larger companies had risky sites in answers about half as often as queries about smaller brands. Questions about regional brand websites triggered risky results more often than other prompt categories. The systems sometimes invented regional sites for brands that did not have them. In one cited conversation, Perplexity named the wrong page as Stüssy’s official website and described it as “the best starting point for guaranteed authentic items.”
The Verge says ChatGPT was the strongest of the three at warning users about risky sites it surfaced, doing so 24 percent of the time. Google spokesperson Jessica Levis Boord said Gemini displays source links to help people verify answers and that Google tests and improves its models. OpenAI spokesperson Taya Christianson said the company is improving safeguards against deceptive shopping results and welcomed researchers’ examples. Perplexity spokesperson Jesse Dwyer said the company does not claim perfect accuracy and emphasized its focus on accuracy. The article does not describe a product change, new access tier or price tied to the test.
What the archive adds
New Runtime's earlier coverage reported that AI shopping agents passed only 57.2 percent of runs in Visa’s benchmark across 22 models and 15 realistic shopping challenges (Visa Agentic Commerce Benchmark, November 7, 2026). The Podqi test is a new fact on the broader shopping-reliability story: it examines the safety of destination links in chatbot answers, rather than task completion in an agent benchmark.
Trend signal
This supports the pattern that agent-ready software exposes capabilities: commerce interfaces need observable, permissioned boundaries, and a surfaced URL should not silently stand in for verified merchant identity. The study tests answers rather than deployed checkout integrations, so it is a warning signal about trust boundaries, not proof that every shopping agent has the same failure rate.
Why now
The Verge frames the results against the coming holiday shopping season, when consumers seek available products and lower prices and may encounter unfamiliar retailers. That makes destination verification more consequential as shoppers use chatbots for discovery and comparison.
Connections to other events
New Runtime's earlier coverage of Visa’s shopping-agent benchmark found a 57.2 percent pass rate across 22 models and 15 shopping challenges (Visa Agentic Commerce Benchmark, November 7, 2026). Together, the findings point to different failure modes in AI commerce: completing a shopping task is one challenge; directing a shopper to a legitimate seller is another.
Why it matters
For a team building shopping assistance, a recommendation is part of the transaction’s security boundary, even if the model never handles payment. A fake or re-registered domain can turn product discovery into a phishing or counterfeit risk, while a dead site wastes a customer’s time. Treating citations as self-authenticating also makes the user carry the burden of distinguishing official stores from imitations.
The practical response is to separate discovery from authorization. Resolve recommended merchants against maintained, brand-specific official-domain data; flag uncertain matches; and require a deliberate verification step before sending users to checkout or asking for sensitive information. Instrument the funnel so teams can measure incorrect destinations, warnings and user overrides. These controls will not guarantee that every merchant record is current, but they make the failure observable and reduce reliance on a fluent answer as a trust signal.
Limits and uncertainty
The findings come from prompts Podqi designed, not a representative sample of shoppers’ conversations. The Verge does not provide the per-model number of answers, confidence intervals, or a breakdown separating recommendations from unannotated displays and warnings; the headline rates combine those behaviors. The test covers 151 brands and two September testing periods, not every category, locale, query style or model version. The study also does not measure whether users clicked the links, lost money, or encountered harm. The Verge did not independently reproduce the experiment, and the company responses do not establish that safeguards have since changed outcomes.
Skill to build now
This week, add a merchant-destination check to one AI shopping flow: compare every suggested URL with a maintained list of official brand domains, label uncertain matches, and block checkout handoff until the user confirms the destination. Log both false positives and missed impersonators so the team can tune the check against real examples.
Reader takeaway
A chatbot can produce a plausible shopping link that is not a trustworthy store. Use the answer to discover candidates, not to authenticate them: verify the domain before a user shares payment or personal information.