Five Poisoned Documents Can Derail AI Search

A Google preprint finds that a handful of query-echoing documents sharply degrades retrieval quality in popular search methods.

Updated on Nov. 5, 2026 in Artificial Intelligence

Five Poisoned Documents Can Derail AI Search

Live Poll

Do you trust AI search tools to give you accurate results?

Google researchers published a preprint on arXiv showing that adding five query-echoing documents per target query cut retrieval quality by 40.2% to 61.5%. The paper, posted November 3, 2026, tested 623 queries drawn from the SciFact and NFCorpus benchmarks.

Why it matters

Retrieval systems underpin many search and AI question-answering tools, so a small number of adversarial documents could skew which sources they surface. The findings also suggest that common defenses can trade protection against attacks for lower quality on ordinary queries.

Across 623 queries from the SciFact and NFCorpus benchmarks, five poisoned documents per target query lowered nDCG@10 by 40.2% to 61.5%. The dense BGE-small retriever showed 100% target exposure under query-only poisoning, while the study has not been peer reviewed.

The details

The attack adds documents to a searchable collection that are written to look highly relevant to a specific query. Full-query anchoring, where a document repeats the full query, makes it a near-perfect match for the BM25 ranking method, which rewards shared words. Overlap-based filters blocked these full-query poisons but also significantly reduced clean retrieval quality, and Maximal Marginal Relevance offered little to no protection in the test setups.

Timeline

  1. November 3, 2026: Preprint posted to arXiv as arXiv:2611.04264.

The Tech Race

The study examines BM25, the long-standing lexical ranking function, and shows its reliance on shared query words makes full-query anchoring an effective attack. The results extend concerns about lexical ranking to dense retrievers such as BGE-small, which showed complete target exposure in the tested setup.

Users of AI search and question-answering tools could see less reliable results if retrieval systems are targeted with poisoned documents. Developers evaluating defenses may need to weigh the tradeoff between blocking attacks and preserving quality on clean queries.

The takeaway

Retrieval systems can be manipulated with very few added documents, so defenses should be tested for their effect on ordinary search quality as well as on attacks. Readers relying on AI search results should be aware that ranking can be influenced by adversarial content.

Further reading

For related coverage, explore more on Artificial Intelligence.

More information

Read the original arXiv preprint on query-relevant corpus poisoning for the full methodology and results.

Live Poll

Do you trust AI search tools to give you accurate results?