Reranking
Reranking is a second, closer look at the passages a search found, putting the best matches first before the AI model reads them.
Reranking is a second pass over search results. A first, fast search pulls in lots of passages that might answer a question. A reranker then reads each one more closely against the question and puts them in a new order. Only the top few go on to the model that writes the answer.
It's where many pages quietly drop out. A page can be found by the first search and still never reach the answer, because the reranker judged other passages a better match.
Why there are two steps
Reading every page closely would be too slow and too costly. So systems split the job. The first search is quick and broad, often comparing the meaning of the question with the meaning of each passage, and it returns a long list. The reranker is slower and more precise, and it only looks at that list.
For example, someone asks "Which coffee subscription lets me pause deliveries?" The first search returns 50 passages about coffee subscriptions. Many mention pausing in passing. The reranker reads each one against the question and moves the passages that answer it directly to the top. A Fernhill Coffee FAQ that says "You can pause or skip any delivery from your account" ranks higher than a Kestrel Roasters page that says pausing is "flexible."
What helps a passage survive
A reranker rewards passages that answer the question plainly and on their own. So:
- Put the answer in the first sentence of the section
- Use the words people use when they ask
- Give facts, like numbers, names and conditions, rather than vague claims
- Keep each section to one question, so it doesn't get diluted
In AnswerOrbit
We can't see the passages a platform dropped. We can show the pages it used. Open an answer in Answers to see the pages it used, with the ones it linked to marked as cited. If the same competitor page keeps being used for a prompt and yours doesn't, compare the two with Page checks in Recommendations.