Re-ranking

Re-ranking is the step in a search or retrieval pipeline where a shortlist of candidate results is scored again, more carefully, and put in a new order. A fast first stage finds candidates; a slower, more accurate re-ranker decides which of them deserve the top spots or a place in an AI answer.

How it works

Scoring every document precisely against every query is too expensive, so systems work in stages. First, a fast method such as BM25 or vector search with cosine similarity retrieves the top few dozen candidates. Then a re-ranker scores each query and candidate pair together. Common re-rankers are cross-encoders, transformer models that read the query and passage at the same time, and LLM-based scoring. Hosted options include Cohere Rerank and Google’s Vertex AI ranking API.

Re-ranking can also blend in other signals: freshness, authority, diversity (for example MMR), or the combined ranks from several retrieval methods, known as fusion.

How it’s used in SEO and marketing

• Understanding AI search. In MLforSEO’s five-stage retrieval funnel, re-ranking is stage three, where deeper relevance, claim-level authority, information density and cross-source consistency start to matter.
• Building your own tools. Retrieve candidate internal links or related posts with embeddings, then re-rank the shortlist with a cross-encoder or an LLM for precision.
• Site search and RAG assistants, where a re-ranker improves which passages the model reads.

Example

In the AI Search & Agentic SEO course, early filtering returns a top K of roughly twenty to fifty chunks by cosine similarity. Re-ranking, a second, more computationally expensive pass using cross-encoders or LLM scoring, cuts that to three to eight. The generator then reads those and typically cites one to four. A chunk that scraped through the first filter on topic alone can lose here if it is vague.

Limitations and things to watch

• Not every production system has a separate re-ranking step, and some use a lightweight model. Treat pipeline descriptions as typical patterns.
• Re-rankers only see what the first stage retrieved. Content that fails early filtering never reaches them.
• Cross-encoders are slow, so they are applied to small shortlists only.
• Search engines do not publish their re-ranking signals, so optimise for clear, specific, verifiable content rather than guessed factors.

Learn it on MLforSEO

• How modern search actually works: the five-stage retrieval funnel
• Indexed is not selected: your AI search problem isn’t ranking
• AI Search Optimisation & Agentic SEO course, lesson on how systems narrow choices.
• Semantic ML-enabled Keyword Research course, lesson on implicit user feedback.