Definition
Hybrid search is a retrieval approach that runs lexical (keyword) search and vector search on the same query and merges the two result lists into one ranking. It exists because each method fails where the other succeeds. Keyword search, usually scored with BM25, is precise on exact strings such as product names, error codes, part numbers and rare terms, but misses paraphrases. Vector search over embeddings matches meaning and synonyms, but can overlook the one exact identifier that matters.
Anthropic's 2024 write-up on Contextual Retrieval gives the standard example: for a query like 'Error code TS-999', an embedding model may return content about error codes in general, while BM25 finds the document containing that exact string. Its tests found that combining embeddings with BM25 retrieved better than embeddings on their own.
The merge step is usually Reciprocal Rank Fusion (RRF), introduced by Gordon Cormack, Charles Clarke and Stefan Büttcher at SIGIR 2009. RRF ignores raw scores, which are not comparable between BM25 and vector similarity, and scores each document by its positions: the sum of 1 / (k + rank) across all input lists, with k set to 60. In their TREC experiments RRF beat the best individual system and the Condorcet fusion method by 4% to 5% on average (Cormack et al.). Elasticsearch's rrf retriever still defaults to a rank constant of 60. Some systems instead combine normalized scores with weights, which allows tuning but needs careful calibration.
After fusion, the merged list typically goes to a reranking model for precise ordering. Perplexity described exactly this design for the search stack behind its answer engine in September 2025: 'We do not force a choice between lexical and semantic retrieval. Rather, we query the search index via both modalities and merge the results into a hybrid candidate set', followed by filtering and progressively stronger rankers ending in cross-encoders (Perplexity). It also criticized vector databases that 'couldn't adequately support even simple boolean filters' and keyword engines with weak semantic ranking, which is the gap hybrid systems close.
Hybrid search is widely available in search infrastructure. Elasticsearch documents it as a standard BM25 query and a kNN vector query fused with RRF, and Postgres setups pair built-in full-text search with the pgvector extension. The standard retrieval-augmented generation pipeline Anthropic describes uses the same combination.
For GEO, hybrid retrieval means a page has two ways to become a candidate, and the best pages qualify through both. Use the exact names people and models search for, including product names, model numbers, plan names and category terms, consistently across pages, so lexical retrieval matches them. Write clear, explicit sentences that state what something is and does, so semantic retrieval places the passage near the questions it answers. A page that relies only on clever phrasing loses the lexical match; a page stuffed with keywords but vague in meaning loses the semantic one and then fails the reranker.
Examples of Hybrid Search
- A support search for 'pump shows E-204 after descaling' matches the exact code E-204 through BM25 and the descaling context through vector search, and RRF puts the one article covering both at the top.
- An electronics retailer's search handles both 'SKU 4471-B' and 'quiet dishwasher for small kitchens' well only after it adds vector retrieval alongside its existing keyword index.
- A RAG pipeline takes the top 50 results from BM25 and the top 50 from vector search, fuses them with RRF, and reranks the merged list before sending 8 passages to the model.
- A B2B software company renames its plans consistently across pricing, docs and comparison pages after noticing AI answers describe its 'Growth' plan using an old plan name found on outdated pages.
