Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Perplexity returned risky shops in 15% of shopping answers, ChatGPT in 8%

Ask a chatbot where to buy a favorite brand's hoodie, and it may name a counterfeit site or an abandoned address. Podqi analyzed 6,795 answers about 151 brands and found such a "risky" site in 15% of Perplexity answers, 8% of ChatGPT answers and 4% of Gemini answers. The count included both recommendations and mentions with a warning, so these figures do not represent the share of customers actually defrauded. Check any link from an AI answer before entering card details, as you would with any ad.

AI-processed from Habr AI; edited by Hamidun News
Perplexity returned risky shops in 15% of shopping answers, ChatGPT in 8%
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Riskyfake shops, old brand domains, dead pages
Key metric #2ChatGPT warned about the risk in 24% of answers that mentioned such sites
Key metric #3Small brands had twice as many risky sites as large ones

Imagine asking a chatbot where to buy a Stüssy hoodie or Birkenstock sneakers, and it confidently names an "official store" that either copies someone else's site or has been abandoned for years. Those are the cases that Podqi, a company that protects brands' intellectual property, set out to measure. It asked ChatGPT, Gemini and Perplexity questions about 151 brands and analyzed 6,795 answers.

A "risky" site was defined as an imitation site or a suspected seller of counterfeits, a former brand domain that someone else has bought, or a non-working page. Such addresses appeared in 15% of Perplexity answers, 8% of ChatGPT answers and 4% of Gemini answers. In other words, roughly one in seven Perplexity answers contained at least one dubious address.

But that is not the share of defrauded buyers: the count included direct recommendations, neutral mentions and warnings about danger alike.

What exactly was checked

Podqi selected 151 brands, from small labels to large manufacturers such as Lululemon, Birkenstock, Stüssy and Le Creuset. Among them was Madhappy, a Podqi client, which is worth keeping in mind when weighing the study. In September 2026, analysts ran two rounds of queries asking the services where to buy the brands' products, which shops to choose, whether international delivery was available, and which regional sites existed. Podqi wrote the queries itself; real buyer conversations were not used. The answers were collected with a tool for tracking AI search results.

The last category of "risky" addresses is debatable: a dead site is not fraud in itself. But a freed-up domain can be bought by anyone, and if a chatbot keeps sending people there, the link leads to an entirely different resource.

Figures by service

Perplexity showed risky addresses most often (15%), followed by ChatGPT (8%) and Gemini (4%). The picture changes when you look at warnings: ChatGPT was the most likely to alert the user to the risk, doing so in 24% of the answers where such sites came up.

For that reason, Perplexity's 15% should not be read as "every seventh buyer ends up with scammers." Mentioning a suspicious address with a warning and recommending a purchase there are different situations. In addition, the results apply to this specific set of queries and to the September test, not to all shopping requests made to these services.

The bots did worst on regional questions. They sometimes listed "official" sites for individual countries even for brands that have no such sites. In one example, Perplexity took a third-party page for the official Stüssy store and called it a suitable place to buy guaranteed genuine products. Smaller brands proved more vulnerable: in answers about large companies, risky sites appeared about half as often. This is logical: there is less reliable data online about a small label, so models more often rely on random pages.

What the companies say and what to do

Podqi CEO Trevor West told The Verge that people should check the source of a recommendation before entering personal data or card details, since a shop appearing in a bot's answer guarantees nothing. Google said Gemini showed risky sites less often than the others, and that source links next to answers help users verify information. OpenAI said it is improving its safeguards and thanked researchers for the examples of errors. Perplexity pointed to strong benchmark results, called accuracy a priority, but stressed that it does not promise complete reliability.

The practical takeaway differs depending on who is reading.

  • For shoppers: open a site from a bot's answer through the brand's official website or its social media accounts, check the domain's age and reviews, especially before sales, when unfamiliar sellers' tempting offers are most persuasive.
  • For brands: monitor what chatbots say about you and anchor official regional addresses on your own website. Don't abandon old domains without a redirect or at least monitoring.
  • For developers of shopping agents: links produced by a model need independent verification, based on a list of official domains, registration age and signs of copied pages.

The study is not flawless. It was conducted by a company with a commercial interest in brand protection, and that company set its own criteria for a "risky" site. But the mechanism behind the problem is realistic: a language model is good at summarizing what it finds, not at guaranteeing that what it finds is real. Until AI shopping gets built-in seller verification, the final check falls to the human.

Story

Key Questions Answered

What exactly was checked?

Podqi selected 151 brands, from small labels to large manufacturers such as Lululemon, Birkenstock, Stüssy and Le Creuset. Among them was Madhappy, a Podqi client, which is worth keeping in mind when weighing the study. In September 2026, analysts ran two rounds of queries asking the services where to buy the brands' products, which shops to choose, whether international delivery was available, and which regional sites existed. Podqi wrote the queries itself; real buyer conversations were not used.

What are the key details on Figures by service?

Perplexity showed risky addresses most often (15%), followed by ChatGPT (8%) and Gemini (4%). The picture changes when you look at warnings: ChatGPT was the most likely to alert the user to the risk, doing so in 24% of the answers where such sites came up. For that reason, Perplexity's 15% should not be read as "every seventh buyer ends up with scammers." Mentioning a suspicious address with a warning and recommending a purchase there are different situations.

What the companies say and what to do?

Podqi CEO Trevor West told The Verge that people should check the source of a recommendation before entering personal data or card details, since a shop appearing in a bot's answer guarantees nothing. Google said Gemini showed risky sites less often than the others, and that source links next to answers help users verify information. OpenAI said it is improving its safeguards and thanked researchers for the examples of errors.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…