Polluted Web Pages Fool LLM Recommenders, New Study Finds

A new paper titled "One Polluted Page Is Enough" reveals that search-augmented LLM recommenders are highly susceptible to web content manipulated by Generative Engine Optimization (GEO). Researchers introduced a benchmark called FORGE, which simulates fake product recommendations, finding that even a single manipulated page can lead to LLMs recommending non-existent products up to 27% of the time. This vulnerability increases when models lack prior knowledge, and built-in reasoning or defenses like skepticism prompts and consensus filters are largely ineffective. AI

IMPACT Highlights a critical vulnerability in LLM-based recommendation systems, potentially impacting e-commerce and consumer trust.

RANK_REASON The cluster contains an academic paper detailing a new benchmark and findings on LLM vulnerabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

• EMNLP 2026 Findings
• Generative Engine Optimization
• Hugging Face
• LLM
• One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →