Google study finds five fake documents can wreck AI search retrieval

A Google study found that adding just five fake documents to a retrieval corpus can substantially degrade AI search accuracy through simple query-echoing attacks, where fabricated text repeats the user's query to game relevance ranking. The research challenges the reliability of retrieval-augmented generation, the method behind many AI search and crypto information tools that pull from live sources before answering. The finding matters because AI-driven search is increasingly used to surface token, protocol and market data, where a handful of spoofed documents could distort what users and automated agents see. Watch for retrieval providers and AI search builders to adopt adversarial filtering, source reputation checks or redundancy requirements to harden pipelines against low-cost corpus poisoning.