AI Chatbots Recommend Fake Retail Websites, Study Finds

What Happened

A study testing shopping-related queries across Gemini, ChatGPT, and Perplexity found that all three systems recommended fabricated retail websites, including domains that did not resolve to legitimate merchants and, in some cases, appeared designed to mimic real storefronts. The test set covered product-specific and category-level purchase queries, and the failure mode was consistent across vendors rather than isolated to a single model or retrieval configuration. The Verge reported the findings, noting that recommended URLs in several cases led to non-existent or misrepresented retail destinations rather than the merchants the models named in their responses.

Why It Matters

Citation grounding — the linkage between a model's stated recommendation and a verifiable, correct source — is the load-bearing assumption behind every shopping, procurement, and recommendation agent currently in production or pilot. When that linkage fails, the failure is not a quality degradation but a liability event: the operator, not the model vendor, owns the downstream harm when a user is routed to a fraudulent or non-existent destination. Retail and commerce agents are among the highest-value deployments for AI builders precisely because they sit close to transaction flow, which means trust failures here carry direct revenue and legal exposure rather than reputational noise. The study also matters because it isolates a class of failure that standard evals miss: models can produce fluent, confident, well-formatted recommendations while the underlying retrieval and attribution layer is silently wrong. Operators who have treated vendor-side citation as sufficient verification now have evidence that it is not, and the burden of grounding shifts to the deployment layer.

Technical Details

The failure mode is consistent with retrieval-augmented and browsing-enabled systems returning generated or hallucinated URLs that are syntactically plausible but semantically unverified — the model produces a domain string matching the merchant's name rather than retrieving the canonical URL from an authoritative index. Fabricated domains passing through recommendation pipelines indicate that post-retrieval validation (DNS resolution, merchant registry lookup, allowlist matching) is either absent or not enforced at the point of response generation. The cross-vendor consistency across Gemini, ChatGPT, and Perplexity suggests the issue is architectural to current browsing and citation stacks rather than model-specific, which limits the value of switching vendors as a mitigation. Precision on the study's query count and per-vendor failure rates is not established in the reporting; operators should treat the direction of the finding as reliable and the specific rates as indicative pending the underlying dataset. The relevant benchmark gap is that citation accuracy is not currently a first-class metric in most published model evaluations, so operators cannot select vendors on grounding reliability using existing vendor disclosures.