Enki

Podqi finds chatbots sending shoppers to risky sites

By

Published

Reporting from The Verge

More onOpenAIGoogle

A Podqi study found Perplexity, ChatGPT and Gemini showed risky or fake retail sites in 4% to 15% of shopping answers.

What it means for founders

If you sell online, a mention is worth little if the answer sends the buyer to a lookalike. Run your own audit: ask ChatGPT, Gemini and Perplexity where to buy your product and what your regional sites are, then log every domain that comes back.

Close the regional gap. Over a third (36%) of the risky sites in regional answers carried a country code or region in the address, Podqi found. Publish one crawlable page listing every official store and name the countries without a regional store. Keep renewing old domains.

If you build agent checkout, treat the model's confidence as untrusted input. As Startup Fortune points out, a person reading a bad link still gets a chance to spot it, while an agent paying through something like ChatGPT Wallet does not. The outlet suggests checking merchants against a vetted list, plus domain age, before money moves. Watch whether claimed brand profiles, which OpenAI now lets businesses set up, start feeding official store data into answers.

What we don't know yet

Podqi sells takedown services, so it has a commercial stake in the finding. The test covered 151 brands over one month with set prompts, so the rates are a snapshot. We have not seen per-engine results for the second round or how often shoppers clicked through.

The story

Perplexity surfaced risky retail links in 15% of shopping answers, against 8% for ChatGPT and 4% for Gemini, in a study by brand protection firm Podqi that The Verge first reported on November 12. Overall, about 9% of 6,795 answers showed a site Podqi rated risky.

How Podqi ran the test

Podqi put the same five shopping questions, covering 151 brands, to the logged-out consumer editions of all three chatbots on matching days in two September rounds. Gauge, an AI search visibility vendor, collected the answers. Podqi then reviewed all 4,437 cited domains.

A site counted as risky if it seemed to impersonate a brand, sold suspected counterfeits, or was dead. Podqi flagged 718 domains: 351 fakes and 367 dead addresses. The dead ones matter because, Podqi notes, a stranger can pick up any lapsed domain and reopen it as a storefront.

What the chatbots did with bad links

When an answer included a risky site, it pitched that site as somewhere to shop 57% of the time, named it without comment 34% of the time and warned against it in 9% of cases. Podqi puts that at 6.3 recommendations for every warning. ChatGPT warned most often, in 24% of answers where a risky site appeared.

Questions about whether a brand ships to your country and what its regional websites are carried the most risk, at 16%. Some answers listed country stores for brands with no official shop there. Smaller brands drew risky links about twice as often as large ones, according to Startup Fortune and TechTimes Vietnam. Enterprise brands got a warning in just 4 out of 125 risky answers. Startup Fortune reports Perplexity presented a wrong page as Stüssy's official site.

Risk fell between the two rounds, from 12% to 7%.

How the companies responded

The companies' replies appear in Daily Pakistan, TechTimes Vietnam and DigitalToday, which summarise The Verge's report. Google conceded Gemini can be inaccurate, noted its lowest exposure and pointed to its source links (Daily Pakistan, DigitalToday). OpenAI said it is tightening safeguards against deceptive shopping sites and, per TechTimes Vietnam, that its systems can still get things wrong. Perplexity said it keeps working on accuracy (DigitalToday) and that no AI can promise full correctness (TechTimes Vietnam).

Sources

Tools in this story

We may earn a commission if you sign up through our links. It never affects our ratings or which stories we cover.

Perplexity logo
Perplexity

The answer engine you can cite

8.8Editor’s scoreVisit Perplexity
ChatGPT

OpenAI's all-purpose AI assistant

9.2Editor’s scoreVisit ChatGPT
Gemini

Google's assistant, built into Workspace and Android

8.9Editor’s scoreVisit Gemini

More in Research

How Enki covers newsCorrectionsReport an error