Reranking · Re-scoring by a relevance model
Reranking is a second ranking pass: after a first retrieval of candidates, a more precise model, often a cross-encoder, reorders the passages by their relevance to the query. It decides what actually reaches the model’s context.
Standalone definition · 36 words
Where does it sit in the pipeline?
Retrieval (BM25, embeddings, RRF fusion) quickly produces dozens of candidates. The reranker scores each query-passage pair more finely and keeps only the best. Metehan Yeşilyurt details this mechanism in his article on rerankers for GEO and AEO (August 2026).
What do his measurements show?
He observed that a BGE-reranker-large places roughly 24% of its attention on punctuation and special tokens for a typical passage, which shows that the form of the text matters as much as vocabulary. He also tested Google’s Vertex AI Ranking API on SEO queries.
What should you take away for content?
A passage should answer the question it addresses directly, without depending on neighboring context, to pass the sort. These are observations on open models: the rerankers of mainstream engines are not public.
Frequently asked questions
Does reranking explain why a well-ranked page is not cited?
It is a plausible explanation: the page may pass first retrieval then lose at reranking. Engines do not publish these steps.
Can you optimize for it?
Mostly by writing standalone, precise passages. No technique is guaranteed.
Related terms
Further reading
- Profound vs Otterly vs Ahrefs Brand Radar: query fan-out comparison
- ← Back to the GEO glossary (114 terms)