Context
Peer-reviewed FORGE research tested 12 commercial and open-weights LLMs using frozen retrieval bundles for 225 products across 15 categories and five shopping scenarios. Locally replacing one retrieved page induced fake-product recommendations at rates up to 27%; replacing the top three pages raised the maximum reported fooled rate to 73.8%.
What changed
The FORGE benchmark isolates the effects of fake product mentions in locally rewritten web evidence, measuring downstream recommendation failures and four imperfect defenses.
Why it matters
A single retrieved source can contaminate generated shopping recommendations, so source admission, source provenance and product-identity validation are distinct from citation visibility.
Visibility OS interpretation
Visibility OS interpretation: a cited product mention is not proof that a product exists or is appropriate. Recommendation systems need evidence validation beyond surface-level source agreement.
What remains uncertain
Do results replicate in live retrieval and across more markets and languages? Which verification approaches reduce fake-product endorsements without suppressing legitimate products?