Re-ranking

Re-ranking is a second-pass scoring step in RAG retrieval systems that refines the initial set of retrieved documents before they’re injected into the LLM’s context. After an initial retrieval (which uses fast embedding similarity or keyword search to find candidate chunks), a re-ranker applies a more sophisticated relevance model to reorder or filter those candidates. The final ranked list determines which sources the AI uses — and cites — in its response.

The retrieval pipeline with re-ranking

User query
    ↓
Initial retrieval — fast, coarse (embedding similarity / BM25)
Returns top-50 candidates
    ↓
Re-ranker — slower, precise (cross-encoder model)
Scores each candidate against the full query
    ↓
Top-5 to 10 re-ranked results
    ↓
Injected into LLM context for generation
    ↓
Response with citations (from top re-ranked sources)

The initial retrieval casts a wide net; the re-ranker selects which documents from that net are actually most relevant.

Why re-ranking matters for brand visibility

Being retrieved is not enough. A page that passes initial retrieval but ranks poorly in the re-ranking step may not make it into the final context window — and won’t be cited. This is why some brands occasionally show up in retrieval audits (we can see your page was fetched) but not in the generated response.

Re-ranking models evaluate factors like:

• Query-document relevance — how specifically the document answers the query (not just topic overlap)
• Content quality signals — fluency, factual density, structural clarity
• Freshness — recently indexed documents may score higher for time-sensitive queries
• Source authority — some re-rankers incorporate domain authority as a signal

What you can do about re-ranking

You can’t configure a retrieval system’s re-ranker directly. But you can write content that scores well on the factors re-rankers typically evaluate:

1. Answer the query directly — re-rankers favor documents that respond to the specific question, not documents that merely discuss the topic broadly
2. Front-load specificity — put the direct answer early; re-rankers evaluate how quickly a document gets to the point
3. Avoid topic drift — a page that starts discussing project management but drifts into general productivity tips will score lower for project management queries
4. Match natural query language — re-rankers compare the query against the document in a cross-encoder; content that mirrors how users phrase questions tends to score higher

Re-ranking and citation order

In responses that cite multiple sources, re-ranking score often influences citation order. The highest-ranked source in the re-ranker’s output tends to be cited most prominently — which is why first-cited sources in AI responses correlate with position drift and share-of-voice metrics. If you appear in citations but consistently near the end of the list, re-ranking performance may be the limiting factor.