Shopping Chatbots Surfaced Risky Sites in Brand Queries, Study Finds

ChatGPT, Gemini and Perplexity each linked to or displayed risky shopping websites in answers to brand-related purchase queries, according to a study by IP protection firm Podqi that was shared exclusively with The Verge on 12 November 2026.

Podqi prompted the three chatbots with shopping questions about 151 brands, both small and large. The test ran over two periods in September. The firm analyzed 6,795 AI answers and classed a site as risky if it fell into one of three categories: fake sites impersonating a real retailer or selling knockoffs, old brand domains re-registered by someone else, or inactive dead sites that could be bought and reused.

Perplexity performed worst, with risky sites appearing in 15 percent of its answers. ChatGPT did so in 8 percent, and Google's Gemini in 4 percent, the lowest of the three. ChatGPT also warned users about the risky sites it surfaced in 24 percent of cases, the highest warning rate among the three engines.

Results varied by query type. Questions about larger companies produced risky sites about half as often as questions about smaller brands. Prompts about regional brand websites produced risky results more often than any other category the study tracked.

The mechanics are not hard to picture. A model that retrieves and cites live web pages inherits whatever its index contains. A parked or lapsed domain that still carries a brand name, or a storefront built to mimic one, can look like a good answer to a retrieval system that weighs relevance over provenance, meaning where a page came from and whether it can be trusted. Smaller brands have fewer authoritative pages, press mentions and long domain histories, so a lookalike faces less competition for the top slot. Regional sites, with unusual domain structures and smaller footprints, may fall into the same gap. The categories suggest the problem is less that the model invents a store and more that it selects one a human would need to vet.

The company describes itself as offering AI-native detection, evidence gathering and automated enforcement across marketplaces, domains, app stores and social platforms. Its website lists protection for trademarks, imagery, product listings, domains, app store violations and advertisements. It says it covers more than 180 platforms, including Amazon, TikTok Shop, Taobao, DHgate, AliExpress, Shein, Shopify storefronts, Google and Meta, and that takedowns typically resolve within 48 hours depending on the platform. It also offers a free audit of IP exposure across those platforms. Trevor West is CEO and cofounder.

In my view, the more durable issue for brand owners is the one the study exposes most clearly: the shopping answer has become a point of sale, yet the trust placed in it has not been matched by the checking a shopper would apply to a search result. A brand that has never looked at what chatbots say about its own domains has no baseline against which to notice a change. Monitoring prompts the way brands once monitored search rankings, and tracking which domains the engines cite, is a sensible starting point.

The broader picture also includes progress under way. Each engine surfaced risky sites in a minority of answers, and one warned users about them in roughly a quarter of cases. Retrieval quality, domain provenance signals and warning behavior are tractable engineering and policy problems. The open question is whether they get solved before shopping agents, which act on an answer rather than simply displaying it, become common. The study measured answers, not transactions. That distinction will matter more as the interface moves from suggestion to checkout.