Safegram app logoSafegram

The AI Found You a Shop. But Is the Shop Real?

An AI shopping recommendation can look authoritative without verifying the seller. A September test, reported in November, exposes the gap—and some practical safeguards.

By Safegram Editorial TeamPublished · 🇮🇪 English
Two adults in an Irish home consider a laptop and parcel while checking a retailer

Written by Safegram Editorial Team

Pause to establish who operates a shop before paying. AI-generated editorial illustration.

Key takeaways

  • The study measured links displayed in tested AI answers, not actual customer losses.
  • A dead or impersonating domain is not the same thing as a proven payment fraud.
  • Confirm the retailer independently before using any payment link suggested by a chatbot.

What the new test actually found

A shopping answer from an AI assistant may sound decisive: here is the best store, here is the regional site, here is where to buy. But a link inside a fluent answer is still a claim to verify. In research publicised by The Verge on 12 November 2026, brand-protection company Podqi examined 6,795 shopping answers produced in September by ChatGPT, Gemini and Perplexity. The researchers asked five repeated questions about 151 brands, rather than observing a random sample of real shoppers.

Across that controlled sample, 9% of answers displayed at least one website Podqi classified as risky. The rates varied: 15% for Perplexity, 8% for ChatGPT and 4% for Gemini. In answers containing a risky website, 57% recommended the address as somewhere to shop, 34% merely mentioned it and 9% warned about it. These proportions describe the tested answers, not a universal error rate or a prediction of how often shoppers will lose money.

Independent boutique owner reviews a laptop among handmade stock

A genuine-looking presentation is not independent proof of a merchant. AI-generated editorial illustration.

“Risky” includes sites that cannot sell anything

Podqi individually examined 4,437 distinct domains and marked 718 as risky: 351 as fake and 367 as dead or inactive. The latter are not necessarily active scams. An inactive site may simply lead nowhere, although a lapsed domain could be registered later by an impersonator. The headline number therefore combines different problems: misleading recommendations, potential brand impersonation and dead links.

The researchers work in brand protection, a business with an interest in identifying these risks. Their prompts, brands and two testing waves were chosen for this project; the study did not measure clicks, completed purchases, financial losses or how the tools perform for every language or user. It is evidence of a specific link-verification problem, not proof that an entire assistant is unsafe or that every named site is malicious.

Why a convincing answer can send you to the wrong place

The highest-risk question in the study asked whether a brand ships internationally and where its regional websites are: 16% of those answers displayed a risky site. A fabricated-looking local shop can resemble a legitimate country outlet. The model may retrieve an old link, confuse a reseller with the brand or present an unverified address with the same confidence as a correct one.

The three tested assistants also showed different strengths. In Podqi’s sample, Gemini surfaced risky websites least often. ChatGPT was more likely to attach a specific warning when it did show one: it warned in 24% of its own risky answers. This is a fairer comparison than declaring one product always safe or another always dangerous. Models and search results change; a September benchmark is not a current guarantee.

Two adults discuss a buying decision in a Dublin café

An extra check can interrupt a rushed buying decision. AI-generated editorial illustration.

A five-minute check before checkout

Use an AI answer to compare product characteristics or build a shortlist, then open the brand’s independently located official home page. Follow the official page’s own regional-store or stockist links, rather than trusting a country-specific address supplied in chat. Check the precise domain spelling, the operator’s identity, business address, return policy and independent reviews. A padlock means an encrypted connection, not that the seller is genuine.

If an unknown retailer demands a bank transfer, sends you to a different payment domain, advertises implausible discounts or pushes an immediate deadline, stop. For an Irish shopper, the Competition and Consumer Protection Commission (CCPC) specifically recommends checking contact information and reviews and avoiding unfamiliar links in social adverts. A hypothetical example: an AI answer names a “brand-ireland” site; verify its existence from the brand’s known main website before buying. That example is illustrative, not a documented incident.

Ireland’s consumer protections do not verify AI links for you

The CCPC’s April 2026 guidance explains that a .ie domain does not by itself establish that a retailer is genuinely Irish. EU consumer rules can provide rights when dealing with a trader, but tracing or enforcing a claim against a dishonest or overseas seller can still be difficult. No Irish, EU, UK or US law makes every AI-recommended store automatically legitimate.

If you have entered payment details on a suspicious site, contact your bank or card issuer promptly; a chargeback may be available depending on the transaction. In Ireland, report suspected fraud to An Garda Síochána and misleading consumer practices to the CCPC. Do not send more identity documents to a stranger simply to “verify” your order.

Craft seller presents a ceramic bowl to a customer at a market

Local sellers benefit when customers can establish identity and provenance. AI-generated editorial illustration.

What brands and AI providers should improve

Brands can make their real country stores and authorised dealers easy to find on their official domains, and monitor lookalike addresses. AI services can distinguish independently verified merchants from unverified links, flag uncertainty and keep warnings next to the precise address involved—not just offer a generic sentence about internet scams.

The lesson for anyone building local commerce or AI-assisted purchasing is broader than one benchmark: convenience is useful only when buyers can inspect who is selling, what evidence supports a recommendation and where a payment goes. Trust requires a verifiable chain from discovery to checkout, rather than an authoritative tone.

Questions readers ask

Does the study mean 9% of all AI shopping purchases are scams?

No. It found risky links in 9% of 6,795 controlled AI answers. It measured neither purchases nor losses, and some classified domains were merely inactive.

Which assistant had fewer risky links in the test?

Gemini showed risky domains in 4% of tested answers, versus 8% for ChatGPT and 15% for Perplexity. These September 2026 figures are sample-specific and may change.

How can I check an “official” local store suggested by AI?

Go independently to the brand’s known global website and follow its country-store or authorised-retailer links. Confirm the domain, contact details, returns and independent reviews.

What if I already paid a fake shop in Ireland?

Contact your bank or card issuer immediately and ask about freezing the card or disputing the payment. Report suspected fraud to An Garda Síochána; the CCPC also accepts misleading-trader reports.

What to do next

Before following an AI shopping link, open the seller’s real home page independently, confirm the destination and protect your payment details. Use AI for research, not as the sole proof of a shop’s identity.

Sources

More from Safegram

Try Safegram

Privacy-first social and a verified marketplace, built in Dublin.