How AI answers are made
Reranking
Also called: Re-ranking
Reranking is a second sorting step in a retrieval system: it first grabs many possible passages quickly, then a more careful model re-orders them and keeps only the best few.
In plain words
The first search step is fast but rough. It can return hundreds of passages of mixed quality.
A reranking model scores each of those passages against the question. Only the top ones go to the language model.
Example
In Anthropic’s own retrieval example, the first step returned the top 150 passages, and the reranker passed on the top 20.
Why it matters for AI search
Anthropic calls reranking a commonly used filtering technique. It gives better answers and lowers cost, because the model reads less.
For your page, being found is not enough. The passage also has to come out as clearly relevant to the question.
What you can do
- Make the relevant passage obviously about the question: name the topic in the heading and the first sentence.
- Keep sections focused instead of mixing several topics.
- Back up claims with specific facts, so each passage has substance.
Often confused with
- Passage ranking
- Passage ranking is a named Google Search system. Reranking is a general step that many retrieval systems use.
Sources
Facts and sources last checked on 19 October 2026.
Related terms
How does your page do in AI search?
GEO-Score checks one page on 22 signals, from AI crawler access to structured data, and shows what to fix first.
Analyze a page