Definition
Reranking is a second-stage retrieval process that takes an initial list of candidate documents—usually returned quickly by vector search or keyword search—and reorders them by relevance using a more accurate, computationally heavier model. The goal is to put the truly best passages at the top before they are passed to a language model, because most RAG systems only feed the top few results into the context window.
The typical pipeline has two stages. First, a fast retriever scans a large corpus and returns a broad candidate set (often dozens or hundreds of chunks) optimized for recall. Then a reranker—commonly a cross-encoder that jointly examines the query and each candidate—scores relevance with much higher precision and reorders the list. This two-stage design balances speed and accuracy: the retriever is cheap and broad, the reranker is expensive but precise and only runs on the shortlist.
Reranking matters because first-stage retrieval often surfaces topically related but not directly useful passages. A cross-encoder can capture nuance—negation, specificity, and intent—that embedding similarity misses, materially improving answer quality and reducing hallucinations. Hybrid systems frequently combine keyword search, vector search, and reranking together.
For GEO, reranking is part of how AI systems decide which sources to actually use. Content that directly and specifically answers a query—rather than loosely matching its keywords—is more likely to survive the rerank stage and be cited, reinforcing the value of answer-first, self-contained passages.
Examples of Reranking
- A RAG pipeline retrieves 100 candidate chunks with vector search, then a cross-encoder reranker selects the 5 most relevant to send to the LLM, sharply improving answer accuracy.
- A documentation search tool adds a reranking stage and finds that pages which precisely answer the query now outrank pages that merely mention the keywords.
- An enterprise assistant combines keyword retrieval, embedding retrieval, and a reranker so both exact-match and semantically similar passages compete for the top slots.
- A GEO team studies reranking by checking which of its passages survive to the top of retrieval for target queries, then rewriting weaker pages into direct, self-contained answers.
