BM25 · Lexical search (sparse retrieval)
BM25 is a lexical ranking function that scores a document by the frequency of the query’s terms, weighted by their rarity and the document’s length. It is the basis of sparse retrieval, often combined with vector search in hybrid search systems.
Standalone definition · 41 words
How does it work?
BM25 counts the occurrences of the query’s words in each document, gives more weight to rare words and limits the effect of repetition and very long texts. It does not understand meaning: “car” and “automobile” remain distinct terms, hence its contrast with dense retrieval.
Why does it matter in GEO?
RAG systems often combine BM25 and embeddings, then merge the rankings, for example with Reciprocal Rank Fusion. Metehan Yeşilyurt studies this chain (BM25, RRF, rerankers) in his experiments on retrieval by AI engines. The query’s exact terms therefore keep some weight, notably for proper names, codes and references.
What should you do in practice?
Keep precise terms and named entities in titles and opening paragraphs, without keyword stuffing, while covering nearby phrasings. Lexical coverage and semantic coverage complement each other.
Frequently asked questions
Has BM25 been superseded by embeddings?
No. It remains strong for exact terms and rare queries, and many systems pair it with embeddings rather than opposing them.
Does Google use BM25?
Google does not publish its index algorithm. BM25 is a reference method in information retrieval, present in many engines and search databases.
Related terms
Further reading
- Profound vs Otterly vs Ahrefs Brand Radar: query fan-out comparison
- ← Back to the GEO glossary (114 terms)