KEY FINDINGS
Four numbers from 6,795 AI shopping answers
9%
of AI shopping answers showed the shopper a risky website.
6.3x
How much more often engines recommended a risky site than warned against it.
16%
of answers to “What are their regional websites?” showed a risky site, the riskiest question.
4 of 125
risky answers about enterprise brands warned the shopper. The other 121 didn’t.
Out of every 100 AI shopping answers
About 9 put a risky site in front of the shopper. Hover the legend to see what the engines did with them.
EXPLORE THE DATA
Where the risk concentrates
This study, conducted by IP protection firm Podqi with data from AI search company Gauge and shared exclusively with The Verge, asked ChatGPT, Gemini and Perplexity the same five shopping questions about 151 brands, on the same days, then reviewed every website their answers named. A site counts as risky if it appeared to impersonate a brand or offer suspected counterfeits, or if it was dead (nothing a shopper could buy from). The charts below show where risky sites appeared, and what the answers said about them.
Results varied by engine
Share of each engine’s answers that showed a risky site. Same prompts, same brands, same days.
Perplexity
15%
ChatGPT
8%
Gemini
4%
Source: Podqi, Sept. 2026 · 2,265 answers per engine
WHAT IT SHOWS
Each engine showed real strengths: ChatGPT warned in 24% of its risky answers, Gemini showed a risky site least often, and Google AI Overviews warned in all 7 of its risky answers. Encouragingly, risk dropped between our two waves, from 12% to 7%. This is fixable.
What engines did with a risky site
Of the answers that showed a risky site, how the engine described it: as a place to buy, without comment, or with a warning.
Recommended it as a place to buy
57%
Mentioned it without comment
34%
Warned the shopper away
9%
Source: Podqi, Sept. 2026 · Mixed answers count as not warned
WHAT IT SHOWS
Recommendations outnumbered warnings 6.3 to 1. The engines know counterfeits exist: one in five mishandled answers warned about them in general. What’s hard is spotting the specific fake in front of them, and that takes ownership data.
Risky answers by question
Select a question to see how often its answers showed a risky site.
RISKY ANSWERS FOR THIS QUESTION
11%
Q1
11%
Q2
6–7%
Q3
6–7%
Q4
6–7%
Q5
16%
Lists are risky by design: every site on a “top websites” list is presented as a place to buy. Two-thirds of risky answers to this question recommended the risky site.
Ranges (6–7%) are as reported in the paper. Source: Podqi, Sept. 2026 · 1,359 answers per question.
WHAT IT MEANS
Analysis and implications
Shoppers used to judge ten blue links. Now an AI answer picks the store for them, so the stakes on getting it right are higher.
01
This is fixable.
Results varied by engine, and we saw real strengths, like ChatGPT’s warning rate and Gemini’s low exposure. Encouragingly, risk dropped between our two waves. Brands, engines and protection vendors each hold part of the answer.
02
Generic warnings don’t protect shoppers.
A warning only helps when it’s attached to the specific site a shopper is about to visit. Cautioning about counterfeits in general while linking a lookalike puts the risk back on the shopper.
03
Showing up in AI answers isn’t enough for brands.
Brands invest heavily to appear in AI answers, but few check where those answers send shoppers. A wrong recommendation becomes a lost sale, a support ticket, a safety concern or a reputational hit, and size is no shield: risky answers about big brands almost never carried a warning.
04
Today’s dead domains are tomorrow’s fakes.
367 of the 718 risky domains were dead. Anyone can re-register a lapsed domain, so an address AI keeps recommending could become a counterfeit storefront.
05
Cross-border shopping raises the stakes.
The riskiest question we tested was the most innocent one: does this brand ship to my country? With cross-border online sales on track to reach $1.2 trillion in 2026, shoppers will ask it more, not less. Treat any AI-supplied regional or “official” URL as unverified, and reach country stores through the brand’s main site.
FOR BRANDS
How to protect your brand in AI search
Three moves any brand can make now, based on where the study found risky sites slipping in.
Publish an official-sites page
Give AI engines one correct list to copy.
List every official regional store and authorized retailer on one crawlable page.
Say plainly which countries have no regional store.
Link it from your footer and help center, and update it when markets change.
Monitor ccTLD and regional variants
36% of risky sites in regional answers used a country ending or region name.
Inventory the country-code domains you own (.co.uk, .de, .com.au) and the ones you don’t.
Watch region and add-on variants, like us-yourbrand.com or yourbrand-shop.com.
Track lapsed brand domains. A dead domain can be re-registered by anyone.
Track your brand’s AI answers
Engines recommended risky sites 6.3x as often as they warned against them.
Each month, ask ChatGPT, Gemini and Perplexity the riskiest questions: regional websites, top websites, where else to buy.
Log every domain the answers surface and flag any you don’t operate.
Note whether each site is recommended, mentioned or warned against.
METHODOLOGY
How we tested
We tracked five shopping questions across 151 brands and three AI engines, over two waves of fieldwork in September 2026, with data from AI search company Gauge. Every domain the engines referenced was reviewed individually.
151
Brands
5
Questions
3
AI engines: ChatGPT, Gemini, Perplexity
6,795
Answers analyzed
4,437
Domains reviewed
718
Risky domains: 351 fake, 367 dead
Data collection
Domain verdicts
Classifying answers
Google AI Overviews
Wave comparison
What the data does not show
WHAT TO DO NEXT
Podqi protects your brand in AI answers
AI search is the new front door to shopping, and impersonators have followed shoppers there. Podqi finds lookalike sites across the web, documents the evidence and takes them down, so there are fewer for AI answers to surface. Brands can close this gap today.
Find fake and lookalike stores
Podqi scans 1,500+ marketplaces, apps and social platforms for impersonating websites, including the regional and “official” lookalikes AI engines surfaced in this study.
Take them down, fast
Set your rules once and Podqi files, tracks and verifies takedowns at the source, with registrars, hosts, marketplaces and platforms.
Keep court-ready evidence
Every infringement is captured, timestamped and packaged with full chain of custody, so your legal team can act on it.
Brandsweep is Podqi’s free, public brand protection tool. Run a scan in minutes, then talk to us about protecting your brand in AI answers.
Questions, Answered.
Everything you need to know before your first takedown.
