On 19 November the Chinese tech outlet 36Kr ran an alert about language models pushing fake shopping sites. The test behind the story came from Podqi, a company that works on brand and IP protection. It sent identical shopping prompts to three chatbots and logged every domain that came back.
The Verge reported the same findings on 12 November. Neither OpenAI, Google nor Perplexity disputes the pattern. All three say they are working on accuracy and advise users to inspect citations and domain names before clicking.
What the test measured
The setup is simple enough to repeat. 151 consumer brands, fifteen questions per brand per platform, 2,265 answers per model. That gives 6,795 responses in total, each one checked for the domains it recommended.
Podqi classified 718 of those domains as risky: 351 were counterfeit or fraudulent storefronts, 367 were dead or expired domains, some of them likely to be picked up by whoever registers them next. Because some answers contained more than one bad link, the number of affected answers is lower than the number of domains.
Per-platform numbers
| Chatbot | Answers with a risky link | Rough count out of 2,265 |
|---|---|---|
| Perplexity | 15% | ~340 |
| ChatGPT | 8% | ~181 |
| Google Gemini | 4% | ~91 |
| Average | ~9% | ~612 |
The spread is wider than the averages suggest. Perplexity is close to four times more likely to surface a bad domain than Gemini in this dataset. That fits how the products differ: Perplexity is built around live web results and cites heavily, so whatever sits in its search index can reach the answer.
Endorsement came far more often than warning
The more damaging finding is not the link itself but the framing around it. In 57% of cases where a risky site appeared, the model endorsed it as somewhere to shop. Risky domains were recommended 6.3 times more often than they were flagged with any warning.
ChatGPT attached a warning in only 24% of the instances where it surfaced a risky link. Read that the other way: three times out of four, a user got a fake storefront with no signal that anything was wrong.
A link in a chat answer carries a borrowed authority. The user asked an assistant, the assistant answered, and the cheap leather bag at the end of the paragraph looks like a recommendation rather than a search result. That is a very different situation from a sponsored listing on a marketplace, where the commercial relationship is at least labelled.
What a 9% rate looks like over ten queries
Nine percent per answer sounds small until you stack queries. If a shopper asks ten separate product questions and each answer has an independent 9% chance of carrying a risky link, the chance of hitting at least one is about 61%. On Perplexity's 15%, ten queries give roughly an 80% chance of at least one bad link. On Gemini's 4% it drops to about 34%.
Those are my own calculations from the published rates, assuming independence, which is an approximation: real users cluster around the same popular brands and the same phrasings, so the practical exposure will vary by person.
The cost of checking a domain
Verification is not an expensive operation. A domain check means a DNS lookup, a comparison against a list of expired or known-fraud hosts, and a check that the domain matches the brand's registered official store. That runs in milliseconds on infrastructure the providers already operate.
Put it next to token prices. GPT-6.2 Astra bills $40 per million output tokens and GPT-6.2 Sol $10 per million. A 300-word shopping answer is roughly 400 output tokens, so a cent or two at most. The check that could have caught a counterfeit shop costs a fraction of the answer that contained it. Cheap is not the same as easy, and the hard part is the brand-to-domain mapping rather than the lookup, but the economics are not the obstacle.
Smaller brands were exposed twice as often
Podqi reports that less digitally canonical brands saw around twice the exposure to fraudulent links compared with major brands. The mechanism is not mysterious: the big names have years of official pages, structured data and retailer agreements behind them, so a model has more correct signal to copy. A mid-sized European manufacturer with a thin web presence gives the model less to work with, and the gap gets filled from whatever ranks.
Where European rules stop
Since 2 August 2026, Article 50 of the AI Act requires operators of conversational systems to tell users they are interacting with an AI, and requires machine-readable labelling of synthetic media. It does not require anyone to verify that a recommended URL belongs to the brand it claims.
The same date brought the European AI Office active supervisory powers over general-purpose models, with fines up to €15 million or 3% of global turnover. Those powers target copyright policy, training-data summaries and systemic-risk documentation. A hallucinated shop link is not on that list.
The high-risk obligations that were once due in August 2026 are now postponed by Regulation (EU) 2026/1744: Annex III systems move to 2 December 2027, embedded Annex I products to 2 August 2028. A shopping assistant was never Annex III in the first place, so the revised calendar changes nothing for consumers here. What remains is national consumer-protection law, which generally assumes a trader on the other side. A dead domain registered by nobody has no trader to pursue.
What a shopper can do about it
Treat a storefront link in a chat answer as unverified. Open the brand's own site and follow its official store locator instead of the URL in the reply. Check whether the domain matches the brand name exactly, and be suspicious of hyphenated variants and country-code domains the brand does not use.
This is not a reason to stop asking models for product categories or comparisons. It is a reason to stop using them as a source of destination links. Our own benchmarks in AI Arena measure tokens per second and time to first token, and those numbers say nothing about whether a URL resolves to the real shop.
Did the test cover only English-language queries?
Podqi evaluated 151 brands with a fixed prompt set and did not publish a per-language or per-market breakdown that I could verify. The 9% average therefore describes the tested set, not European retail as a whole, and the real rate for a Czech or German shopper asking about local retailers is unknown.
Does reporting a fake link to the chatbot help?
It feeds into the provider's feedback loop, but none of the three companies has published figures on how quickly a reported domain stops appearing. The study found warnings in a minority of cases even before any user report, so reporting is a slow correction rather than a fix.
Is a dead domain as dangerous as a counterfeit shop?
Not immediately, but it is a supply of addresses. An expired domain that a brand once used can be registered by someone else and turned into a lookalike store, and the model may keep pointing at it because the URL appears in older citations.
The study measured 151 brands with a fixed prompt set over several rounds in September 2026. Podqi has not released the full domain list or a breakdown by market, which limits any attempt to estimate the European share of the 718 flagged addresses.