Reranking

How does reranking work?#

Reranking works by taking the top results from a fast first-stage search and re-scoring each one with a slower, more accurate model that reads the query and the passage together, then reordering the list so only the best-matching few reach the answer.

The split exists because of a speed-accuracy trade-off. First-stage retrieval — keyword matching and vector embeddings — compares representations computed in advance, fast enough for billions of passages. A reranker, usually a cross-encoder, runs the query and the passage through the model together, which is far more accurate and far too slow to run on everything. The Sentence-BERT paper puts numbers on the gap: finding the most similar pair among 10,000 sentences takes about 65 hours with BERT reading every pair, and about 5 seconds with precomputed embeddings.

The Sentence-Transformers documentation describes the standard recipe: retrieve the top 100 or so candidates with a fast bi-encoder, then re-rank those hits with a cross-encoder. The approach took off in 2019, when Nogueira and Cho used BERT as a passage re-ranker and beat the previous best result on the MS MARCO benchmark by 27% (relative, MRR@10).

Do AI search engines use reranking?#

AI search engines rerank before they answer: Perplexity documents cross-encoder rerankers as its final ranking stage, Anthropic’s published retrieval setup reranks 150 candidates down to 20, and every engine that fans out many searches must cut the pooled results to a handful of citations.

Perplexity’s architecture write-up is the most explicit: fast lexical and embedding scorers narrow the set, then cross-encoder rerankers make the final cut at document and sub-document level, and the rankers keep learning from signals across millions of user requests an hour. Google and OpenAI both describe query fan-out — several searches per question — which produces a pool far larger than the few sources an answer cites.

Anthropic sums up why the step exists: reranking ensures “only the most relevant chunks are passed to the model,” which “provides better responses and reduces cost and latency because the model is processing less information.” For a site owner, that means the model never reads most of what was retrieved — only what survived the rerank.

How do you optimize content for reranking?#

Optimize content for reranking by writing passages a model would judge a direct answer when it reads them next to the question: the question’s key terms restated, the answer in the first sentence, specific facts, and nothing that depends on text outside the passage.

1. Restate the question’s subject and terms. A cross-encoder reads the query and passage together, so a passage that says “Plannora pricing for a team of five” plainly addresses a question about exactly that.
2. Answer in the first sentence. The reranker is judging whether this passage answers this question; make that obvious from the first line. See answer-first content.
3. Be specific. Numbers with units, names, dates and conditions separate an answer from a description.
4. Stay on one question. A passage covering three topics is a partial match for all three.
5. Make it self-contained. Rerankers see the passage, not your page. See content chunking.
6. Keep it current and dated. Perplexity prefilters stale content before its rerankers even run. See content freshness.

Reranking vs ranking: what's the difference?#

Ranking orders an entire index for a query, while reranking re-scores only the short list a first ranking produced, using a more expensive model — so ranking decides whether you’re a candidate and reranking decides whether you’re chosen.

In classic search the ranked list is the product. In AI search it’s an intermediate step: a page can rank on page one for a sub-query and still lose at reranking to a passage that answers the question more directly. Citation lists rarely match the classic top 10 — only 37.9% of pages cited in Google AI Overviews rank top 10 for the typed query, per Ahrefs’ March 2026 data. Ahrefs attributes that mainly to fan-out; reranking is the step that then picks among the pooled sub-query results.

• Ranking gets you retrieved: crawlable, indexed and matching the sub-query. This is classic SEO, and it’s the entry ticket.
• Reranking gets you chosen: the passage that best answers the question, judged next to the question.
• Grounding gets you cited: the model attaches a citation to the passage it actually used. See grounding.

Related terms#

• RAGHow LLMs answerRetrieval-augmented generationA technique in which an AI system first retrieves relevant documents from an index and then gives them to a large language model to write its answer, so the response can cite current sources instead of relying only on what the model memorized in training.Read the entry
• How LLMs answerSemantic searchA retrieval method that matches a query to content by meaning and intent — typically by comparing vector embeddings — instead of by the exact words both contain, so a page can be found for phrasings it never uses.Read the entry
• How LLMs answerVector embeddingsNumerical representations of text — lists of hundreds or thousands of numbers — arranged so that passages with similar meaning sit close together, which lets search systems match a query to content by meaning rather than by shared keywords.Read the entry
• How LLMs answerContent chunkingThe splitting of a page into smaller passages — often a heading and the text beneath it — that an AI retrieval system indexes, scores and quotes individually, which makes the section, not the whole page, the unit that competes for a citation.Read the entry
• How LLMs answerQuery fan-outAn AI search technique in which the engine rewrites one user question into several narrower sub-queries, runs them in parallel and builds its answer from the combined results — which is why a page can be cited for a prompt it doesn’t rank for.Read the entry
• Content & relevanceAnswer-first contentA writing structure that puts the direct answer to a section’s question in its opening sentence and adds context and evidence after it, so readers and AI retrieval systems can take the answer without reading further.Read the entry