- Podqi tested three chatbots on 151 brands. Risky sites were shown or mentioned, including in answers that warned about them, in 4% of Gemini, 8% of ChatGPT and 15% of Perplexity answers.
- The prompts were Podqi's own. The Verge reports that smaller brands appear especially at risk, and chatbot answers can surface lookalike or dead sites.
- Ask what chatbots say about your brand, check old and regional domains, and tell customers how to verify a store.
A brand's address book now sits inside a chatbot
A company controls its website, its listings and its ads. It does not control what a chatbot says when someone asks where to buy its products. A new study by Podqi, an intellectual property protection firm, shows why that gap matters.
Podqi shared the work exclusively with The Verge. The lesson here is that a brand's identity now lives partly in answers it does not write and cannot see.
What Podqi tested
Podqi asked ChatGPT, Gemini and Perplexity shopping questions about 151 brands, large and small. The prompts included "What are the top websites for buying [brand]?" and "Does [brand] ship internationally? What are their regional websites?" The brands included Lululemon, Birkenstock, Stüssy, Madhappy and Le Creuset. Madhappy is a Podqi client.
The test ran in two periods in September and produced 6,795 answers. The prompts were not taken from real user conversations. Podqi wrote them itself and ran them through a tool that tracks AI answers across chatbots.
What counts as risky
Podqi defined risky sites in three ways. The first is a fake site that impersonates a real retailer or sells knockoffs. The second is an old brand domain that someone else has registered again. The third is a dead site that is no longer active but that someone else could buy and reuse.
The numbers
Of the three chatbots, Gemini had the lowest rate. Podqi counted a risky link or mention in 4 percent of its answers. ChatGPT's rate was 8 percent. Perplexity's was 15 percent.
A caveat applies. Podqi counted any mention. That covers recommending a risky site, showing it without comment, and warning the user about it. So not every counted answer pointed a shopper to a bad site.
Warnings varied. Podqi said ChatGPT warned about the risky sites it surfaced 24 percent of the time. That was the highest warning rate of the three.
How a chatbot ends up citing a bad site
The following is general background, not a finding from Podqi. The Verge's report does not say why the chatbots cited these sites, and the chatbots' internal checks are not public.
Some chatbots that answer shopping questions look up web pages, then write a reply and attach links. This is a simplified description of retrieval-based answers, not a description of how these three products work. In such a design, the lookup matches the wording of the question against pages it can find. Matching on wording does not by itself tell the system who owns a site today. OpenAI says it is adding safeguards.
That creates two weak points. An old domain that someone re-registered may still carry the brand's name and a trail of links from years ago. To a lookup, it can still look relevant. A dead domain that anyone can buy is the same problem waiting to happen.
Regional questions add a third. Asked for a brand's regional websites, a chatbot is being asked for a list. If the brand has no such sites, a system built to produce a helpful reply may fill the list with plausible-looking addresses. This is an interpretation, and Podqi's test did not examine the cause.
Smaller brands and regional sites
Answers about larger companies contained risky sites roughly half as often, Podqi found. The Verge reads this as a sign that smaller brands appear especially at risk.
One plausible reason is that a lesser-known retailer leaves less reliable information for a chatbot to draw on. The source does not confirm that.
Regional questions drew risky sites more often than other prompt types. The chatbots also sometimes listed fake regional sites for brands that have none. For Stüssy, Perplexity presented an incorrect page as the brand's official site. One reading is that a brand with no presence in a country leaves a gap for lookalikes to fill.
What the companies said
Podqi's chief executive, Trevor West, said consumers should verify a source before entering personal or card details. A Google spokesperson said Gemini shows source links so people can check information. An OpenAI spokesperson said the company is continuing to improve safeguards against deceptive sites in shopping results.
Perplexity's spokesperson, Jesse Dwyer, pointed to the company's record on research benchmarks and said Perplexity does not claim perfect accuracy. The Podqi study measured risky-site citations, not benchmark accuracy.
What to ask your team
This is one study, run by an IP protection firm that counts Madhappy among its clients. Its prompts were its own, and it covered 151 brands. Read it as a reason to look, not as a measure of every chatbot answer.
Start with five questions. What do the main chatbots say when asked where to buy from you? Which regional sites do they name, and do you run them? Which old domains have you let lapse? Who watches for lookalike stores, and how fast can they act? Do customers know how to confirm your official address?
Brands cannot edit a chatbot's answer. They can find out what it says before a customer does.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: The Verge.
