- Query planning: the prompt is rewritten and split into several searches — see What is a query fan-out.
- Retrieval (wide net): keyword search (BM25) and vector search run together; result lists merge via Reciprocal Rank Fusion, rewarding documents that rank well in both. OpenAI’s documented File Search works this way (800-token chunks, 400 overlap, hybrid search, optional reranker).
- Reranking (fine sieve): a cross-encoder-style model reads each candidate passage against the query and re-scores the shortlist. Your page is often scored as passages, not one URL — a strong page can contain a weak candidate passage, which is why page-level SEO metrics cannot explain every citation win or loss.
- Generation: the LLM writes from the surviving passages and may cite some.
How AI search chooses passages — retrieval, reranking, generation
Peec’s reranker guide (Aug 2026) — AI answers run query fan-out, hybrid retrieval merged with RRF, a reranker sieve, then generation; each stage fails differently, so diagnose the stage first.
AI search picks sources through a pipeline, and each stage fails differently. Peec’s reranker guide (full guide, Aug 2026) describes the common documented pattern:
