top of page

ChatGPT Shopping Risks: Study Flags Fake Retail Sites

20 minutes ago
11 min read
Illustrative photograph. Photo: Sanket Mishra / Pexels.

ChatGPT shopping risks surfaced in a Podqi study that found risky retail links in answers from ChatGPT, Gemini, and Perplexity. Across 6,795 answers, 9% included a website the firm classified as risky. That category covered suspected impersonation or counterfeiting sites and inactive domains.

The brand protection vendor collected responses through Gauge, an AI search company, using five shopping questions about 151 brands. Its AI shopping study examined logged-out consumer services during two September 2026 waves. Researchers reviewed 4,437 domains and classified 718 as risky: 351 suspected fake sites and 367 inactive ones.

The findings expose a specific tension: finding a plausible shopping destination and verifying its operator are different jobs. An assistant can answer a product question fluently while leaving the seller’s identity unresolved. For shoppers, brands, and chatbot developers, that gap matters precisely when research turns into a purchase decision.

What the ChatGPT shopping risks study actually measured

Podqi reported risky links in 15% of Perplexity answers, 8% of ChatGPT answers, and 4% of Gemini answers. Among answers containing risky sites, 57% recommended them, 34% mentioned them without comment, and 9% warned users. Those percentages describe the risky subset, not every shopping answer.

The finding concerns destinations and recommendations, not a measured rate of completed shopping fraud.

That distinction separates several questions readers might otherwise collapse together. Did an answer contain an unsuitable destination? Did its wording encourage a visit? Did anyone follow the link, enter payment information, or lose money?

The study’s headline figures address the earlier stages. They should not become a claim that the same share of shoppers experienced fraud. Exposure, persuasion, purchase, and loss are separate outcomes.

An inactive address also presents a different immediate problem from an operating impersonation store. Both undermine a shopping answer, but they require different descriptions. Calling every flagged destination an active scam would exaggerate the evidence.

Conversely, describing the issue as ordinary broken links would overlook the suspected impersonation findings. The useful interpretation preserves both categories. A recommendation system needs working destinations and credible evidence about who controls them.

The sampling context matters for the comparisons, too. These results describe the tested consumer experiences, prompts, brands, and collection period. They do not establish a permanent safety ranking across every configuration or shopping feature.

For example, a general chatbot answer and a dedicated merchant integration should be evaluated separately. A test of one experience cannot automatically certify or condemn another. That boundary matters when readers compare familiar product names with unfamiliar implementations.

The study therefore offers a concrete starting point for scrutiny rather than a universal scorecard. It directs attention toward the transition between answering a shopping question and endorsing a seller. That transition deserves evaluation in its own right.

A useful reading asks what evidence would justify the recommendation at that moment. Does the system know the address belongs to the brand? Has it verified an authorization relationship? Or has it merely found language that matches the request?

These are analytical questions raised by the findings, rather than mechanisms the study conclusively establishes. The difference matters. Observing a questionable output does not reveal every internal step that produced it.

The practical consequence is a narrower, more defensible conclusion: shoppers should not treat a conversational retail recommendation as completed seller verification.

A shopping answer carries more weight than a product summary

Product research and seller selection involve different kinds of judgment. An assistant might help someone compare materials, sizes, or product categories without deciding where that person should spend money.

Once it supplies a purchase destination, the task changes. The answer connects a product preference to a business that will receive payment and personal information. Accuracy now includes the identity of the seller.

A relevant product description does not establish a legitimate commercial relationship.

Consider an illustrative shopper looking for a jacket from a particular brand. An assistant can correctly identify the jacket’s features while supplying an unrelated store. The product information and the purchasing destination need separate checks.

The same distinction applies when an answer labels a store “official” or “authorized.” Those words express a relationship, not a resemblance. A familiar logo or matching product photograph cannot establish that relationship by itself.

Conversational presentation makes the distinction especially important. A direct answer can read like a completed judgment, even when the underlying evidence remains incomplete. The user receives a recommendation in the assistant’s voice.

That creates pressure on chatbot developers to clarify what their wording promises. “This page mentions the product” is a limited claim. “Buy here” invites action and requires stronger support.

Brands face another problem: a correct mention of their name can still accompany an incorrect destination. Visibility alone therefore cannot describe the quality of a shopping answer. Where the answer sends someone matters as much as whether the brand appears.

For a retailer, an unauthorized recommendation raises questions about customer confusion and responsibility. A shopper might associate a bad experience with the brand whose name appeared throughout the answer. That is a plausible business risk, not a loss measured by this study.

The implications also reach workplace purchasing. An employee researching equipment or merchandise can carry an unverified seller link into a procurement discussion. A polished summary might then circulate separately from the original conversation.

In that situation, the important record includes the source and its verification status. A shortlist that says where a recommendation came from remains easier to challenge. A bare list of suggested stores hides the unresolved judgment.

For teams maintaining a searchable knowledge base, approved purchasing sources belong alongside the relevant research. A retained link still needs review; saving it does not certify its operator.

The commercial pressure is therefore specific. Assistants that help identify products also need boundaries around seller endorsement. Brands need to inspect the destinations attached to their names. Buyers need a visible distinction between discovery and approval.

Citations establish a source, while seller identity needs separate evidence

A citation can help readers inspect where an answer obtained information. It does not automatically establish that a cited business deserves their money.

That difference becomes clear when separating the claims inside a shopping answer. A page might support a product specification, show an item listing, or describe a return policy. None independently proves authorization by the manufacturer.

Evidence that a page exists is weaker than evidence that its operator is the seller the shopper intended to reach.

This is the central reversal in the story. Features that make answers feel inspectable can coexist with unresolved commercial trust. A link adds something readers can examine, but examination still has to happen.

The risk is especially apparent in an illustrative international shopping request. A user might reasonably expect a brand to operate separate country stores. A regional-looking address can fit that expectation without establishing ownership.

The question then becomes relational: how does this address connect to the known brand? A link from an established brand site would support a different claim from matching words in an address.

This is why asking a chatbot to repeat that a store is legitimate provides limited reassurance. A second assertion is useful only if it introduces better evidence. Confidence and verification should not become interchangeable.

The same principle applies to warnings. A general reminder about online scams leaves the specific recommendation unresolved. A warning attached to a particular destination tells the shopper which action needs reconsideration.

Answer design should preserve that specificity. If the system cannot verify a store, its recommendation language should reflect the uncertainty. Otherwise, the burden moves silently to someone who asked for help choosing.

Consumer protection guidance already distinguishes presentation from legitimacy. The Canadian Centre for Cyber Security’s online shopping guidance cautions against relying on website appearance or padlock icons alone.

That advice matters here because trust can accumulate through several weak signals. An assistant sounds certain, the store looks polished, and the connection appears secure. Together, those details can feel stronger than the underlying identity evidence.

For developers, the analytical challenge is to evaluate the claim the interface actually communicates. A system that retrieves a relevant address has completed a discovery step. It still needs a basis for describing that address as a suitable place to purchase.

This does not mean every recommendation requires identical handling. Established retailers, independent shops, marketplaces, and brand-operated stores have different relationships. The system should describe the relationship it can substantiate.

A cautious answer can remain useful. It can identify a candidate seller while stating what remains unverified. That approach preserves the value of research without presenting uncertainty as a completed endorsement.

The chatbot comparison is useful, but it does not settle AI shopping safety

The differences between ChatGPT, Gemini, and Perplexity deserve attention. They show that the tested services did not produce interchangeable outcomes. They also invite questions about what each comparison actually captures.

An assistant might surface fewer questionable destinations because it chooses different sources, supplies fewer stores, or handles uncertainty differently. Output rates alone cannot fully distinguish those explanations.

A lower observed exposure rate supports a narrower finding than a blanket claim that one chatbot is safe.

Warning behavior adds another dimension. Avoiding a destination and identifying a problem after mentioning it are different capabilities. A useful evaluation should show both, alongside the usefulness of the resulting answer.

Otherwise, an assistant could appear cautious simply by declining shopping questions. That would reduce exposure while providing little help. A shopping product needs to balance useful discovery with justified recommendations.

The commercial source of the research also warrants scrutiny. Podqi sells brand protection services, giving it a business interest in the problem it describes. That interest does not invalidate the observations.

It does make transparent classification important. Independent reviewers would need enough evidence to understand disputed cases. A suspected counterfeit seller, an unrelated reseller, and an unavailable website should not become indistinguishable labels.

Reviewers should also ask how uncertainty was handled. Some retail relationships are easier to verify than others. A methodology needs room for unresolved cases rather than forcing every address into a confident verdict.

There is another boundary between identifying a problem and proving a proposed remedy. A vendor can document questionable recommendations without establishing that its own service fixes every underlying cause.

The same standard applies to chatbot providers. A statement about safety improvements would need evidence tied to shopping destinations. General claims about answer quality would leave this particular issue unresolved.

This article has not independently replicated the experiment. The study should therefore remain attributed throughout, and its classifications should remain the researchers’ classifications. Its findings are a reason to investigate, not permission to overstate certainty.

The comparison also needs time attached to it. An observed outcome belongs to a particular collection period. Later product behavior requires fresh testing rather than assumptions based on a familiar company name.

For buyers asking about AI shopping safety, the most useful answer is consequently conditional. The study identifies reasons for caution across the tested services. It does not establish that every recommendation fails or that every flagged address causes harm.

A meaningful response would make uncertain seller identity easier to see. It would also make contested results easier to reproduce and correct. Those changes would address the specific trust gap exposed here.

Existing shopping safeguards still apply at the AI handoff

The study raises a new interface question around an established consumer problem. An assistant can introduce a seller, but the shopper still needs a basis for trusting that business.

The Federal Trade Commission’s seller verification advice recommends researching unfamiliar businesses and consulting reviews from several sources. It also warns that reviews and ratings can mislead.

A chatbot recommendation should begin a seller check, rather than serve as its final piece of evidence.

That principle does not require treating every unfamiliar store as fraudulent. Small retailers and independent businesses can be legitimate. The relevant question is whether the available evidence supports the specific purchase.

A useful distinction separates uncertainty from accusation. “I cannot verify this seller” communicates a limit. “This seller is fraudulent” alleges wrongdoing and needs stronger evidence.

Assistants should preserve that distinction when responding to follow-up questions. A system that swings from enthusiastic recommendation to unsupported accusation has not solved the verification problem. It has changed the direction of its confidence.

For shoppers, the handoff from conversation to website deserves deliberate attention. The destination address should match the seller being evaluated. Evidence about one business should not be carried over to a similarly named business.

The FBI’s holiday shopping guidance likewise emphasizes checking unfamiliar sellers before a purchase. Its warnings cover problems including paying for goods that never arrive.

Those established safeguards remain relevant regardless of how the destination was discovered. A recommendation from a chatbot does not alter the seller’s obligations or establish its track record. The discovery channel and the merchant are separate parts of the transaction.

For organizations, the issue is less about banning assistants than assigning responsibility clearly. Product research can use conversational tools while seller approval follows an established purchasing process.

The person reviewing a purchase should be able to see whether the store is already approved. If it is unfamiliar, the record should say what was checked. That creates an accountable decision rather than an inherited assumption.

There is also a product design opportunity here. An assistant could distinguish verified merchant relationships from ordinary web suggestions. The wording would need to explain the actual basis for that distinction.

A label alone would not be enough. “Verified” needs a defined scope: verified ownership, verified authorization, or something else. Without that explanation, the label risks adding another impression of certainty.

The policy question follows the same logic. Responsibility becomes harder to assess when an interface mixes research, recommendation, and commercial promotion. Clear boundaries would help users understand which claims they are being asked to trust.

ChatGPT shopping risks therefore concern more than whether a link opens. They concern the evidence and language surrounding a purchasing suggestion. The final handoff should make unresolved trust visible before the user acts.

What to watch next: verification, repeat testing, and real outcomes

The next useful developments will be observable changes in shopping answers and evidence quality. More confident language would not resolve the problem. Better substantiated recommendations would.

Progress means improving the destination, the endorsement, and the evidence available to inspect both.

First, watch for product changes that distinguish merchant discovery from verified relationships. Relevant changes would include clearer seller labels, specific warnings, and explanations of what has actually been checked.

The strongest signal would connect a recommendation to inspectable evidence about the business. A generic caution below an otherwise confident store recommendation would offer weaker support.

This distinction should also appear in correction behavior. If a user reports an unsuitable destination, the response should address that address specifically. Repeating broad shopping advice would leave the original recommendation unresolved.

Second, watch for independent repeat testing with transparent classifications. A reproducible evaluation should separate unavailable destinations, suspected impersonation, suspected counterfeiting, and uncertain cases.

It should also preserve the context needed to interpret comparisons. The tested experience, collection period, and recommendation wording matter. Without those details, a ranking risks appearing more universal than its evidence allows.

Useful replication would test whether improvements hold beyond a familiar set of questions. It should examine whether a system remains helpful while becoming more careful about seller identity.

That is a harder standard than removing every unfamiliar result. It rewards justified assistance instead of silence. It also gives developers a clearer target for improvement.

Third, watch for evidence connecting recommendations to actual shopper outcomes. Clicks, purchase attempts, complaints, and confirmed losses answer different questions. They should remain separate measures.

Outcome evidence would help establish the practical consequences of a questionable recommendation. It could also reveal where interventions matter most. A warning before visiting a store and a warning after entering payment details have different purposes.

Conversely, limited evidence of completed harm would constrain claims about prevalence. That would not make incorrect recommendations acceptable. It would make the description of their consequences more precise.

The response from brands will matter alongside chatbot changes. Clear, accessible information about official stores and authorized relationships would give verification efforts something concrete to inspect.

Still, publishing that information and having assistants use it correctly are separate milestones. The industry needs evidence of the connection between them. Announcing a source of trustworthy information does not prove consistent use.

For readers, the immediate takeaway is a boundary around trust. ChatGPT shopping risks do not erase the value of AI product research. They show why a suggested seller deserves its own evaluation.

Before your next AI-assisted purchase, examine the handoff between recommendation and transaction. Does the answer establish who operates the destination, or does it simply present a plausible store? If the evidence stops at relevance, keep the purchase decision open. Compare the address with information from a trusted brand source, inspect independent seller information, and retain the basis for your decision. For workplace purchases, carry that evidence into the approval process instead of forwarding a bare link. The question worth asking is specific: what supports trusting this business with this transaction? Treat a confident answer as a claim to inspect, then decide whether the evidence earns the recommendation.

Give every agent the context to do better work

Connect your agents to the knowledge, decisions, and history already organized in remio.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page