Research and dataBy 6 min read

AI sends shoppers to fake sites: 9% of "where to buy" answers. What brands can do

Podqi study of 6,795 ChatGPT, Gemini and Perplexity answers: 9% point to fake or dead sites and recommend them 6 times more often than they warn. What to do.

Bar chart: share of answers with risky sites per AI engine in the Podqi study, with Perplexity, ChatGPT and Gemini

In short

A Podqi study of 151 brands and 6,795 answers collected in September 2026 finds that 9% of answers to "where do I buy this brand" point to a risky site: fake, counterfeit or dead. When that happens, the engine recommends the site 57% of the time and warns only 9% of the time. Perplexity is at 15%, ChatGPT at 8%, Gemini at 4%. Your brand can defend itself with an official-sites page, monitoring of look-alike domains and a monthly check of what the engines cite.

  • 9% of ChatGPT, Gemini and Perplexity answers to "where to buy" questions contain a risky site: 718 domains out of 4,437, half fake and half dead.
  • When a risky site appears, the engine recommends it 57% of the time and warns only 9% of the time: recommendations outnumber warnings 6.3 to 1.
  • The regional-websites question is the riskiest (16%); smaller brands are hit more often, but of 125 risky answers about enterprise brands only 4 warned.
  • Netcraft shows that some of these sites are built specifically for AI engines using GEO techniques: they are not just hallucinations.
  • Five-point defense: official-sites page, look-alike domain monitoring, risky questions every month, official feed and profile, reporting of clones.

Written by the Geosnap AI agent, reviewed and approved by Rinald Sefa.

Translated from the Italian original. Read the original

When someone asks ChatGPT, Gemini or Perplexity where to buy a brand, about 9 answers in 100 point to a risky site: a store impersonating the brand, a seller of suspected counterfeits, or a dead domain. That is the finding of a Podqi study covering 151 brands and 6,795 answers collected in September 2026. For anyone selling online this is not only a reputation issue: if the AI sends a shopper to a fake site instead of yours, you lose the sale and the shopper loses the money. Here are the numbers, the riskiest questions, and what to check on your site and in AI answers.

What did the Podqi study measure?

It asked the same five shopping questions about 151 brands to ChatGPT, Gemini and Perplexity, on the same days, in the logged-out consumer versions. The questions are the ones people actually type when they want to buy: the top websites for buying the brand, where to buy similar products, where to find it online, where else to buy it besides the official site, and whether it ships internationally and has regional websites. Gauge ran the collection in two waves in September 2026, for a total of 6,795 answers, 2,265 per engine.

Every domain named in the answers was reviewed one by one: 4,437 domains. A site counted as risky if it appeared to impersonate the brand, offered suspected counterfeits, or was dead, meaning nothing could be bought from it. Mixed answers, which name a site without taking a position, count as "not warned". Podqi is an IP-protection company and Gauge helps brands appear in AI answers: two declared interests to keep in mind, even though the method is described in a way that can be checked.

How often does AI point to a risky site?

In 9% of answers. Of the 4,437 domains reviewed, 718 were risky: 351 fake and 367 dead. The more serious finding is how they are presented: when a risky site appears, the engine recommends it as a place to buy 57% of the time, mentions it without comment 34% of the time, and warns the shopper only 9% of the time. Recommendations outnumber warnings 6.3 to 1. Per 100 answers, on average, 5 recommend a risky site, 3 mention it and 1 warns.

EngineAnswers with at least one risky siteNote
Perplexity15%Highest of the three engines compared
ChatGPT8%Warned the shopper in 24% of its risky answers
Gemini4%Lowest of the three
Google AI Overviews7 risky answers in totalWarned in all 7; sample too small for a percentage

Between the first and second September waves, risk fell from 12% to 7%. That suggests the engines are filtering better, but also that the picture changes week to week: a one-off measurement says little, as we showed in AI answers change every week.

Which questions are riskiest, and for which brands?

The regional-websites question is the worst: 16% of answers to "does the brand ship internationally, what are its regional websites?" contain a risky site, and 36% of the risky sites in those answers use a country ending or a region name in the domain. Next comes "what are the top websites for buying the brand" at 11%, where two thirds of risky answers recommend the site. The other three questions sit between 6% and 7%.

Size is no shield. Of 125 risky answers about enterprise brands, only 4 warned the shopper; the other 121 did not. According to Futurism's summary of the full study, smaller brands were hit more often. The mechanism is simple: the less clear and consistent a brand's online footprint, the more the engine has to guess which domains belong to it.

Hallucinations or targeted attacks?

Both, and the second is growing. Netcraft, a security company that tracks online fraud, ran nearly 3,000 questions about real brands' login pages through ChatGPT, Copilot, Gemini and Perplexity in June 2026: 1.7% of answers contained a malicious link, or 0.28% of the 20,706 links returned across 2,905 answers. In 2025 the same company had found that 34% of the sites cited for a brand were not controlled by that brand.

What has changed is the method. According to Netcraft, the risk has moved from invented or parked domains to attacker-controlled infrastructure: pages built specifically to be retrieved by AI engines, using the same techniques an honest brand uses to get cited. Engines work by retrieval and generation: when they find a new source they must judge it on the fly, with little information. A well-written clone site with structured data and answer-style copy can get through. For the brand the damage is twofold: it loses the citation and sees its name tied to a scam.

What your brand can do, in practice

Make it unambiguous which sites are yours, and check every month what the engines answer. Podqi's and Netcraft's recommendations, combined, become a five-point list.

  1. An "Official sites and authorized retailers" page. List every regional store and every authorized retailer, state the countries where you have none, link it from the footer and the help center, and keep it crawlable by AI bots. It is the source engines look for when someone asks "where do I buy this".
  2. Monitor variants of your domain. Country endings, prefixes and suffixes such as "shop" or "store", and above all old brand domains left to expire: dead sites were 367 of the 718 risky ones.
  3. Ask the risky questions yourself, every month. Ask the engines where to buy your brand and what your regional sites are, log every domain returned and whether it is recommended, mentioned or warned against. In Geosnap you can keep these prompts among those you monitor on ChatGPT, Gemini and Google AI Mode and see which domains get cited instead of yours. Perplexity is coming.
  4. Give the engines first-hand data. For ChatGPT there is an official channel: the product feed in the Agentic Commerce Protocol format, currently open to approved partners who apply at chatgpt.com/merchants. OpenAI does not describe a seller verification process, but reserves the right to remove products or ban sellers. We covered it in ChatGPT Shopping picks from product feeds. In parallel, claim your profile in Businesses in ChatGPT: a verified profile gives the engine an official anchor.
  5. Report and take down. Sites impersonating your brand should be reported to the registrar and the host, as with any phishing case. In AI answers there is currently no form to correct a wrongly recommended site: the only lever is making the official source clearer and easier to find than the copy.

Two cautions on reading the data. The study used logged-out consumer versions and English questions about brands that are presumably international: for a brand selling in a single country the "regional websites" question matters less, while "where can I buy it online" is identical. And the numbers move fast: between two waves in the same month, risk almost halved. What stays is that for shopping questions AI recommends more often than it warns, and the difference between your site and its copy is up to you: a clear official-sites page and a periodic check of what gets cited.

Sources

Frequently asked questions

How often does ChatGPT recommend fake sites when I ask where to buy?

According to Podqi's September 2026 study, 8% of ChatGPT's answers to five shopping questions about 151 brands contained at least one risky site, and ChatGPT warned the shopper in 24% of those cases. Perplexity was at 15% and Gemini at 4%. The average across the three engines is 9%.

What does "risky site" mean in this study?

A domain that appears to impersonate the brand, offers suspected counterfeits, or is dead, meaning nothing can be bought from it. Podqi reviewed the 4,437 cited domains one by one and classified 718 as risky: 351 fake and 367 dead.

My brand sells in a single country: does this affect me?

Yes, although less on the regional-websites question, which is the riskiest in the study. The questions "where can I buy it online" and "what are the top websites for buying it" apply to any brand. Expired brand domains and variants with "shop" or "store" are the first place to look.

Can I ask OpenAI or Google to remove a fake site from the answers?

There is currently no form to correct a wrongly recommended site in AI answers. You can report the clone to the registrar and the host, as with phishing, and make your official source clearer: an official-sites page, a verified profile in Businesses in ChatGPT and, for online stores, a product feed in the format OpenAI requires.

How do I check what the engines answer about my brand?

Ask the study's five questions about your brand yourself, every month, on ChatGPT, Gemini and Google AI Mode, and log every domain returned and how it is presented. A monitoring tool such as Geosnap tracks these prompts over time and shows which domains get cited instead of yours.