You probably check whether ChatGPT names your brand. You probably don't check which address it sends shoppers to.

A study by IP protection firm Podqi checked 6,795 AI shopping answers about 151 brands. About 9 percent put a risky site in front of the shopper. Of the 718 risky domains, 367 were dead, and the question about a brand's regional websites drew risky sites more often than any other.

My read is that part of this is brand housekeeping. Podqi doesn't say who used to own those dead addresses. Old and country domains are still the piece a brand can check for itself.

Podqi published the study on November 12, 2026, and shared it first with The Verge, which ran it as a shopper safety story.

Podqi sells brand protection, which gives it a commercial reason to find fakes. It worked with Gauge, an AI search company, to collect answers from the logged-out consumer versions of ChatGPT, Gemini and Perplexity over two waves in September 2026. Each engine got the same five where-to-buy questions about the same brands on the same days. According to The Verge, Podqi wrote its own prompts and did not use real user conversations.

Then Podqi reviewed every domain the answers named. A site counted as risky if it seemed to impersonate the brand or sell suspected counterfeits, or if it was dead, with nothing a shopper could buy from. The brands ranged from Lululemon and Le Creuset to Madhappy, which The Verge notes is a Podqi client.

Regional questions drew the most risky sites

The question that went wrong most often was the most innocent one: "Does [brand] ship internationally? What are their regional websites?" It showed a risky site in 16 percent of answers. The "top websites" question came next at 11 percent, and the other three landed at 6 to 7 percent.

Podqi found that 36 percent of risky sites in regional answers used a country ending or region name. The Verge reported that the engines at times listed fake regional sites for brands that have none. In one Perplexity answer, the wrong page was named as Stüssy's official website and called "the best starting point for guaranteed authentic items."

Podqi's study page showing a Perplexity answer that labels a page as Stüssy's official site, with that line highlighted
The answer Podqi uses as its example, with the line that sent shoppers to the wrong page. Screenshot: Podqi, Nov. 13, 2026

Of the 718 risky domains, 351 were fake and 367 were dead. The Verge describes dead sites as ones that aren't active but that someone else could buy and reuse. Podqi warns that anyone can re-register a lapsed domain, and an address AI keeps recommending could become a counterfeit storefront.

The engines split, and warnings were rare

Perplexity showed a risky site in 15 percent of its answers, ChatGPT in 8 percent and Gemini in 4 percent, according to The Verge.

Column chart: share of answers showing a risky site, Perplexity 15 percent, ChatGPT 8 percent, Gemini 4 percent
Same prompts, same brands, same days. A check on one engine says little about the other two.

When an answer did show one, it recommended the site as a place to buy 57 percent of the time. It mentioned the site without comment 34 percent of the time and warned the shopper 9 percent of the time. Podqi sums that up as recommendations outnumbering warnings 6.3 to 1.

Of the three engines, ChatGPT warned most often, in 24 percent of its risky answers. Podqi lists Google AI Overviews separately. They warned in all 7 of their risky answers, too few to lean on. Questions about larger brands surfaced risky sites about half as often as questions about smaller ones.

The companies answered The Verge by email. Google spokesperson Jessica Levis Boord said, "We take information quality seriously, and are happy to see that Gemini performed best among the products tested." OpenAI spokesperson Taya Christianson said, "We're continuing to improve our safeguards to make potentially deceptive sites less likely to appear in shopping-related results."

Perplexity spokesperson Jesse Dwyer pushed back on the ranking: "Perplexity is consistently the most accurate AI on credible research benchmarks, and has been for four years."

The study covers one month and one vendor's prompts. Risk also fell from 12 percent in the first wave to 7 percent in the second. I read the engine ranking as a snapshot that may already have moved.

Disclosure: my agency, ViewEngine, sells AI visibility work and builds its own tracking tools.

Where I land

The Verge told this as a shopper story, and that part is real. I care more about what a brand controls. AI answers are built from everything the web says about a brand. That can include addresses it stopped paying for and country stores it never explained.

My guess is that some of those 367 dead domains are old brand, campaign or regional addresses. Podqi doesn't break that out, and it didn't test whether fixing them changes the answers. Still, renewing the domains you let lapse is cheap. So is one crawlable page that lists every official store and says which countries have none, which Podqi recommends too.

The same prompts showed risky sites in 4 percent of Gemini answers and 15 percent of Perplexity answers. Checking one engine tells you little about the others.

When I measure AI visibility, I want two columns. One is how often the brand gets named. The other is where the answer sends people, and I think that second column gets checked far less than it should.

Free newsletter

The Sherpa Brief

One email a week on what changed in AI search, what it means, and what to do about it. Written for marketers, SEOs, and publishers.

Free. Unsubscribe in one click.