Reranking is a second round of evaluation in which the system re-assesses the order of documents found in the first, fast search. The first step is deliberately coarse and cheap – it pulls out dozens to hundreds of candidates. The reranker then assesses each query-passage pair far more precisely and selects the ones that actually make it into the model's context and have a chance of being cited. What decides this is a direct match with the specific sub-query, not the overall authority of the website, which explains why smaller websites appear in AI answers alongside large brands. There is only one way to prepare for this: have a passage that answers the exact question asked directly and in the first sentence, without an introduction and without the need to work out context from another part of the page.
See also: Hybrid search, Content chunking, Grounding.
