LLM Recommenders Vulnerable to Web Content Pollution
August 23, 2026
A study using the FORGE framework shows that a single polluted web page can fool LLM recommenders into promoting fake products at rates up to 27%. Across 12 commercial and open-weights models, search-augmented retrieval is highly susceptible to Generative Engine Optimization (GEO) attacks.
HOW THIS AFFECTS YOU
●
builderYou must implement robust verification layers for RAG-based recommendation engines.
●
policyThis highlights a critical security risk in how AI mediates consumer information.