Perplexity, ChatGPT and Gemini steer shoppers to fake stores

A study by Podqi, a company that protects brands against counterfeiting, found that ChatGPT, Gemini and Perplexity all sent shoppers to risky online stores. According to The Verge, which reported the results on 12 November 2026, and to the summary by Startup Fortune, Perplexity showed a risky retail site in 15% of its answers, ChatGPT in 8% and Gemini in 4%. The findings arrive a few weeks before Black Friday, when many people ask a chatbot where to buy.

How did Podqi run the test?

According to Infobae, Podqi ran its tests in two rounds in September 2026 and asked about 151 brands, among them Lululemon, Birkenstock, Stüssy, Madhappy and Le Creuset. The questions were of the type “what are the best sites to buy this brand?” or “does it have regional sites?”. In total the company analysed 6,795 answers, according to Startup Fortune and Digital Today.

Podqi counts a site as risky if it imitates a real store, sells counterfeits, is an old brand domain that someone else has re-registered, or is a dead page that anyone could buy and reuse, according to Startup Fortune. A domain is simply the web address of a site. Of the 4,437 domains reviewed, 718 were flagged: 351 fake and 367 dead, according to the same outlet. The dead ones matter because an address that nobody runs today can lead to a scammer's page tomorrow.

Which chatbot did worst, and did they warn users?

Share of answers containing a risky shopping site, according to Startup Fortune and Digital Today:

• Perplexity: 15% of its answers.
• ChatGPT: 8% of its answers.
• Gemini: 4% of its answers, the best result of the three.
• All three together: about 9% of the 6,795 answers.

Warnings were rare. According to Digital Today, in 57% of the answers that contained a risky site the chatbot recommended it as a place to buy, in 34% it mentioned the site without any explanation and in only 9% it warned of the risk. ChatGPT warned most often, in 24% of the cases where it displayed such a site, according to Infobae. Startup Fortune calculates that risky sites were recommended 6.3 times as often as they were warned against.

One case shows the pattern. According to Infobae, Perplexity named a wrong page as the official Stüssy site and presented it as a good starting point for authentic goods. Nobody was hacked: the chatbot read the web, picked a page and vouched for it.

Which brands and questions were most exposed?

Smaller brands had roughly double the exposure to risky links of large ones, according to Startup Fortune. Questions about regional storefronts were the riskiest, at 16%, according to Startup Fortune and Digital Today. Part of the reason is that some of those regional sites do not exist: Digital Today reports cases where fake sites were suggested for brands that have no official regional store.

What do Google, OpenAI and Perplexity say, and what are the limits?

According to Infobae, Google spokeswoman Jessica Levis Boord said the company was pleased that Gemini performed best and recalled that it shows source links so users can check answers. OpenAI said its goal is for ChatGPT to connect people with trustworthy places to buy and that it keeps improving its safeguards. Perplexity spokesman Jesse Dwyer said it is consistently the most accurate AI in research tests and never claims to be 100% accurate; he did not explain the Stüssy case.

The figures call for caution. The prompts were written by Podqi rather than taken from real user conversations, according to Infobae. Podqi sells brand protection, and Madhappy, one of the brands analysed, is its client, according to the same outlet. The test covered 151 brands in a single month, so Startup Fortune says the percentages are a snapshot rather than a standing rate. The sources read do not say how many shoppers were actually defrauded.

Is this problem new?

No. According to Infobae, in July the firm Searchable tested more than 72,000 questions to ChatGPT, Gemini and Perplexity about UK businesses: about one answer in 15 containing web addresses pointed to a dead link, a copy or a different business. Perplexity failed in 10% of cases, ChatGPT in 4% and Gemini in 5%. Infobae also cites British firm Ask Silver, which in June 2026 found two cloned Russell & Bromley sites among the sources of a ChatGPT answer; OpenAI then said it had removed them from its search index.

Why does it matter for AI shopping agents?

A shopping agent is software that clicks and pays on the user's behalf. Startup Fortune notes that a chatbot suggesting a bad link leaves a person one click from catching the mistake, whereas an agent that clicks for you does not. It argues that a model's confidence in a retailer cannot serve as a trust signal, and that a list of verified merchants, a domain-age check and a hard stop before payment details go to an unknown domain are cheaper than the first chargeback dispute. Liability when an agent buys from a clone store, it adds, is not settled.

What it changes for a company in Brazil

The sources read do not report results for Brazil or for Portuguese-language questions, so no Brazilian rate can be claimed. The mechanism, however, applies to any brand: the chatbots appear to fill gaps with whatever the web offers. For a Brazilian retailer or brand, the practical steps are these:

• Publish and keep up to date your official stores, including regional or country pages, so that chatbots find a real answer.
• Ask ChatGPT, Gemini and Perplexity where to buy your brand and check which addresses they return.
• Keep control of old domains, because abandoned addresses can be re-registered by others.
• Tell customers to type the brand name in a search engine and enter through an address they already know.

For shoppers, Podqi CEO Trevor West advises verifying any source before entering personal or card data, going straight to the brand's own site, and watching for misspelled domains and odd extensions, as reported by Startup Fortune.