Reranking

Retrieval in an AI answer engine usually happens twice. The first search is fast and broad: it pulls a long list of passages that might be relevant. The second, reranking, is slow and careful: it reads each candidate against the question and puts them in a new order. Only the top of that new order reaches the model that writes the answer.

How does reranking work?

question
→ first-stage retrieval (keywords and/or vectors)   → ~100+ candidates
→ reranker (reads question + passage together)      → new order
→ top few passages                                  → language model
→ answer with citations

The difference between the two stages is how they compare. A first-stage search has to cover millions of documents in milliseconds, so it compares the question and each passage separately: by matching words, as in BM25, or by distance between precomputed vectors, as in vector search. A reranker, usually a cross-encoder, reads the question and a passage together and scores how well that passage answers that question. It is far more accurate and far too slow to run on everything, which is why it only runs on the short list.

What is reranking worth?

Anthropic published a measured example. In its test of contextual retrieval, the share of queries where the right chunk was missing from the top 20 was 5.7% at baseline. Better first-stage retrieval, combining contextual embeddings with BM25, cut that to 2.9%. Adding a reranker, which scored 150 retrieved chunks and kept 20, cut it to 1.9%.

The cost is time. Reranking adds a step to every query. Engines decide how many candidates to rerank by trading accuracy against latency and money, and none of the consumer AI engines publishes that setting.

Can you see reranking in real answers?

Only its outcome, and you should be suspicious of anyone who claims more. What every engine shows is the list of links under its answer, in order. Average a domain's place in that list across many answers and you get the visible trace of retrieval and reranking together.

Here is that trace for the answers to the 61 questions Attensira tracks about its own category, between 4 September and 4 October 2026, read over the Attensira MCP: 100,380 citations in 4,524 answers.

Official sources are linked early: developers.openai.com at a mean position of 2.20 and openai.com at 3.11, with the trade publication searchengineland.com at 2.70. Our own attensira.com sat at 3.31 across 1,525 citations. g2.com reviews, GitHub repositories and LinkedIn posts are cited too, but late, at 4.44, 4.85 and 6.70, as supporting evidence rather than as the answer. That fits what a reranker rewards: a passage that answers the question precisely. It is a pattern in the outcome, not a reading of any engine's scores.

Position is an index into one answer's list of links, averaged per domain. It is not a ranking of a brand, and a move from 3.3 to 3.0 on its own is not evidence of anything.

Why should a publisher care?

Because reranking is where "found" turns into "used". Being in the first-stage results gets your passage onto the short list. The reranker then compares it with the other candidates on one question: which passage answers this best? General, well-ranked pages lose that comparison to specific ones.

What helps:

• One question, one passage. Put the direct answer in the first sentences under a heading that asks the question.
• Self-contained wording. The reranker sees a passage, not the page. Name the subject instead of writing "it" or "this tool".
• Specifics. A number, a limit or a named condition makes a passage a better match for a precise question.

How does Attensira use it?

Attensira cannot see any engine's reranker, and does not pretend to. It records the outcome: every citation in every answer, in order, and ranks the pages that decide each question by citation share times position weight, so a page linked first in many answers counts for more than one linked last in a few. When your page is retrieved but cited late, the crew drafts the passage that answers the question directly, for your approval. See win this answer.