Definition
Reranking is a second-stage retrieval step in which a slower, more accurate model rescores a shortlist of candidate documents or passages and reorders them by relevance to the query. The first stage, usually vector search, keyword search or both, is built to find everything that might be relevant quickly. The reranker then decides which few candidates are actually worth passing to the language model, which matters because most RAG systems put only the top handful of passages into the model's context.
The two stages trade off speed and precision. A first-stage retriever compares a query with millions or billions of items using precomputed vectors or an inverted index, so it cannot read each item closely. A typical reranker is a cross-encoder: it reads the query and one candidate together in a single pass and outputs a relevance score. That joint reading captures details similarity scores miss, such as negation, the exact entity asked about, or whether a passage answers the question or only mentions its keywords. Because it is expensive, it runs only on the shortlist. Newer listwise rerankers score a whole candidate list in one call, and general-purpose LLMs are also used as rerankers.
The gains are large and well documented. In Anthropic's 2024 Contextual Retrieval tests, retrieving the top 150 chunks and reranking them down to 20 brought the top-20 retrieval failure rate from 5.7% to 1.9% when combined with contextual embeddings and BM25. When Voyage AI released its rerank-3 model on September 30, 2026, it reported that reranking improved nDCG@10 by 20.33% on top of plain BM25 results and by 7.65% on top of its own voyage-3-large embeddings, averaged over 95 datasets; rerank-3 accepts documents of up to 32,000 tokens and can follow instructions about what counts as relevant (Voyage AI). Other current options include Cohere's rerank-v4.0-pro and rerank-v4.0-fast, which handle multilingual text and semi-structured JSON (Cohere), and Jina's jina-reranker-v3.5, released in August 2026 with weights published under a non-commercial license.
AI search engines rerank at web scale. Perplexity's description of its search stack, published in September 2025, says early ranking stages use fast lexical and embedding scorers, and 'as the candidate set is gradually winnowed down, we then use more powerful cross-encoder reranker models to perform the final sculpting of the result set'. Google, OpenAI and Microsoft do not publish their ranking stacks, but the pattern of cheap retrieval followed by expensive rescoring is standard in search.
For GEO, reranking is the point where a page that was retrieved still loses. A reranker reads your passage against the exact query and rewards direct, specific answers. Passages that open with the answer, name the entity, include concrete figures and stay on one topic score well. Passages that circle the subject, or match the query's words without answering it, drop out even when the page ranks well in classic search. This is the mechanism behind the advice to write answer-ready content and to make every section of a page stand on its own, as described under sub-document retrieval and passage ranking.
Examples of Reranking
- A RAG pipeline retrieves 100 candidate chunks with hybrid search, a cross-encoder reranker scores each against the question, and only the top 5 are sent to the model.
- For 'Does Plan B include SSO?', vector search returns a pricing overview, an SSO setup guide and a Plan B feature table; the reranker moves the feature table to the top because it answers the question directly.
- A developer docs team adds a reranker with the instruction 'prefer pages for API version 3' and stops getting answers built from deprecated version 2 pages.
- A content team rewrites a vendor comparison so each product section opens with price, seat limit and main integration, and the page starts appearing among the cited sources for comparison prompts.
