Reranking

Reranking is the second stage of a retrieval pipeline — how bi-encoders and cross-encoders reorder retrieved results by relevance before they're served or handed to an LLM, and what that means for AI search visibility.

First published: Jul 3, 2026 · Last updated: Jul 17, 2026 · Advanced
AIAuthored byAI v2

Reranking is the second stage of a retrieval pipeline: a cheap, broad first pass retrieves a candidate set of documents or passages, then a slower, more precise model re-scores and reorders that shortlist before results are served or fed to an LLM. The core mechanic is bi-encoder vs cross-encoder — a bi-encoder encodes the query and document separately into vectors and compares them (fast, scalable, less precise), while a cross-encoder encodes them together and scores the pair directly (slower, more accurate). You can't score a whole billion-page corpus with the expensive model, so you retrieve broadly and rerank the shortlist. Google doesn't use the word 'reranking' publicly, but its named BERT and passage-ranking systems do the job, and Microsoft documents an explicit Bing-derived reranker in Azure AI Search. Reranking is not the same as Reciprocal Rank Fusion. The SEO upshot: because rerankers score query-passage pairs jointly, self-contained, unambiguous passages that read as a direct answer score better.

TL;DR — RerankingReranking is the second stage of a retrieval pipeline: after a cheap, broad first pass pulls a candidate set of documents or passages, a slower, more precise model re-scores and reorders that shortlist by true relevance before the results are served or handed to an LLM. is the second stage of a two-stage (or multi-stage) retrieval pipeline: a cheap, broad retrieval pass (BM25Semantic search is meaning-based retrieval — matching what a user means, not just the words they typed. Search engines detect entities, expand synonyms, infer intent, and rank by conceptual relevance, which is why keyword stuffing lost its power and topical depth gained it. keyword match, embeddingEmbeddings are dense numerical vectors — lists of floating-point numbers — that represent the meaning of text in a high-dimensional space. Semantically similar content lands close together, so search and AI systems can match by meaning, not just keywords./vector similarity, or both) pulls a candidate set, then a slower, more precise model re-scores and reorders that shortlist. The core mechanic is bi-encoder vs cross-encoder — a bi-encoder encodes query and document separately into vectors and compares them (fast, precomputable, less precise); a cross-encoder encodes them together and outputs one relevance score per pair (slower, can’t be precomputed, more accurate). You can’t run a cross-encoder over a whole corpus, so you retrieve broadly and rerank the shortlist. Google doesn’t say “reranking” publicly, but BERT and passage rankingPassage ranking is a Google AI system that scores individual sections ('passages') of a page so a single page can earn multiple relevance scores for different queries. Google still indexes whole pages — only the ranking changed. do the job; Microsoft documents an explicit Bing-derived reranker in Azure AI SearchAI search uses large language models and retrieval-augmented generation (RAG) to synthesize an answer from multiple sources rather than returning a ranked list of links. Examples include Google AI Overviews, ChatGPT Search, and Perplexity.. Reranking ≠ Reciprocal Rank FusionHybrid search runs keyword (lexical/BM25) search and vector (semantic) search together, then merges the two result lists into one ranking — commonly with Reciprocal Rank Fusion — so it catches both exact-term matches and meaning-based matches that either method alone would miss. (RRF). SEO upshot: rerankers score query-passage pairs jointly, so self-contained, unambiguous passages win.

The retrieve-then-rerank pattern

Two-stage retrievalVibe ranking is informal practitioner shorthand for LLM-based reranking — using a large language model to reorder retrieved search candidates by holistic, semantic judgment (relevance, clarity, expertise, completeness) rather than keyword overlap or vector similarity alone. It is not an official Google or Bing term. trades candidate breadth against more expensive scoring. Evidence for this claim A cross-encoder can score query-document pairs for reranking after an initial retrieval stage. Scope: Sentence-BERT evaluation and related retrieve-then-rerank use; cross-encoders are one reranking approach, not a universal implementation. Confidence: high · Verified: Reimers and Gurevych: Sentence-BERT Model choice and latency-quality tradeoffs are implementation-specific. Evidence for this claim A rerank model can reorder an existing candidate list by relevance to a query. Scope: Cohere's Rerank product behavior; inputs, limits, and scoring semantics are vendor-specific. Confidence: high · Verified: Cohere: Rerank overview

Reranking changes the order only after retrieval creates the candidate set. Source: Reranking

A query enters fast first-stage retrieval, which produces a candidate shortlist. A slower query-candidate scoring model reranks only that shortlist into the final order. A document omitted by retrieval never reaches the reranker.

© Patrick Stox LLC · CC BY 4.0 ·

Every large-scale relevance system faces the same problem: you can’t afford to run your most accurate relevance model on your entire corpus. So the standard solution is to split the work into stages. Google Cloud’s own search documentation states the logic plainly: “In short, retrieval is finding relevant documents, while ranking is ordering those retrieved documents. Ranking all the available documents can be computationally expensive. Therefore, retrieval and ranking work sequentially.” (Google Cloud, “About retrieval and ranking”)

Stage one — retrieval — casts a wide net cheaply. It uses lexical matching (BM25 over an inverted indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed.), embedding-based vector searchVector search finds content by comparing the meaning of a query against stored content as numerical vectors, retrieving the closest ones in a high-dimensional embedding space. At scale it uses approximate nearest neighbor (ANN) algorithms — not exact comparison — to search billions of vectors in milliseconds., or a hybrid of the two, and returns a candidate set. Stage two — reranking — takes that shortlist and re-scores every candidate with a more expensive, higher-precision model, then reorders. The one-line version everyone converges on: retrieve cheaply and broadly, rerank precisely on a small set, then serve or generate.

Bi-encoders vs cross-encoders: the core mechanic

The whole topic hinges on one architectural distinction — when the query and the document meet.

  • Bi-encoder (the first-stage retriever). It encodes the query and each document separately, each into its own vector, and then compares the two vectors with something like cosine similarity. Because the document vectors don’t depend on the query, you can compute and index them ahead of time, which is what makes retrieval fast enough to run across an entire corpus. The cost: query and document never actually interact, so the model has to, in effect, compress every possible meaning of a document into a single vector — and nuance gets lost. Bi-encoders are what embeddingsEmbeddings are dense numerical vectors — lists of floating-point numbers — that represent the meaning of text in a high-dimensional space. Semantically similar content lands close together, so search and AI systems can match by meaning, not just keywords. and vector searchVector search finds content by comparing the meaning of a query against stored content as numerical vectors, retrieving the closest ones in a high-dimensional embedding space. At scale it uses approximate nearest neighbor (ANN) algorithms — not exact comparison — to search billions of vectors in milliseconds. are built on.
  • Cross-encoder (the stage-two reranker). It encodes the query and one candidate document together, as a single joint input through a transformer, and outputs a single relevance score for that pair. Because the model sees both at once, it can directly weigh how the specific words of the query relate to the specific words of the document — far more accurate. The cost: nothing can be precomputed. Every query-document pair has to be run through the model at query time, so it’s far too slow to apply to a whole index. That’s precisely why it’s reserved for the shortlist.

Google, notably, describes this exact mechanism in its own words. In the Google Cloud retrieval/ranking docs, one of the listed retrieval signals is cross-attention, defined as something that “allows a model to consider the relationship between a query and a document to assign a relevance score to the document.” That is the cross-encoder idea under a different name.

Why not just use the accurate model on everything?

Latency and cost make it infeasible at scale, and the gap is enormous, not marginal. Pinecone’s write-up on two-stage retrieval puts a concrete number on it: on a 40-million-record set, running a BERT-style cross-encoder reranker over everything on a V100 GPU would take more than 50 hours, versus under 100 milliseconds for vector searchVector search finds content by comparing the meaning of a query against stored content as numerical vectors, retrieving the closest ones in a high-dimensional embedding space. At scale it uses approximate nearest neighbor (ANN) algorithms — not exact comparison — to search billions of vectors in milliseconds.. (Pinecone, “Rerankers and Two-Stage Retrieval”) That’s the entire justification for the two-stage design — you get most of the cross-encoder’s accuracy while only paying its cost on a few dozen or few hundred candidates.

Vectara frames the same myth directly — the question of why not just score all documents with the most precise model if it’s available — and the answer is the same: you can’t, so you filter cheaply first. (Vectara, “What is reranking and why does it matter?”)

How Google does this

Google has never published an official statement using the terms “reranking,” “cross-encoder,” or “bi-encoder” about Google Search itself — worth stating plainly so we don’t overclaim. But the function is documented under other names.

Google’s own Guide to Google Search Ranking Systems names two systems that do reranking’s job:

  • BERT — “an AI system Google uses that allows us to understand how combinations of words express different meanings and intent.” BERT jointly reads the words of a query in context; a BERT-based reranker scores query-document relevance the way a cross-encoder does.
  • Passage rankingPassage ranking is a Google AI system that scores individual sections ('passages') of a page so a single page can earn multiple relevance scores for different queries. Google still indexes whole pages — only the ranking changed. — “an AI system we use to identify individual sections or ‘passages’ of a web page to better understand how relevant a page is to a search.” That’s reranking at the passage level rather than the page level (see passage rankingPassage ranking is a Google AI system that scores individual sections ('passages') of a page so a single page can earn multiple relevance scores for different queries. Google still indexes whole pages — only the ranking changed. for the deep dive).
  • RankBrain — Google’s earlier system that “helps us understand how words are related to concepts,” so it can return relevant content even without exact-match words.

Google Research has also published the mechanism outright: its paper Learning-to-Rank with BERT in TF-Ranking describes encoding queries and documents with BERT and applying a learning-to-rank layer on top, and explicitly frames it as passage re-ranking — reporting the best performance on the MS MARCO passage re-ranking task as of March 30, 2020. That’s a Google Research publication rather than Search Central product guidance, so treat it as Google’s technical research, not a statement about the live Search pipeline.

One number worth hedging: the “cut down to the top 1,000 results, then reorder them” framing that circulates widely in SEO traces back to my own conference deck’s interpretation of public research and patents — not a current, verbatim Google statement about web Search. Google Cloud’s enterprise search product does document a concrete pipeline (“the model retrieves documents in the order of thousands… The ranking model then orders the retrieved documents and serves the top 400 ranked results”), but that’s the Vertex AI SearchAI search uses large language models and retrieval-augmented generation (RAG) to synthesize an answer from multiple sources rather than returning a ranked list of links. Examples include Google AI Overviews, ChatGPT Search, and Perplexity. product, not Google web Search. Don’t assume either the 1,000 or the 400 applies to Google Search itself.

How Bing/Microsoft does this

Microsoft is much more explicit, and its clearest documentation is the closest thing to an official production-reranker description you’ll find. Azure AI Search’s semantic ranker is documented as “a feature that measurably improves search relevance by using Microsoft’s language understanding models to rerank search results” — and crucially, “the underlying technology is from Bing and Microsoft Research.”

The mechanics map cleanly onto the two-stage pattern:

  • It “always adds secondary ranking over an initial result set that was scored using BM25 or Reciprocal Rank Fusion (RRF).” Stage one is BM25 or RRF; the semantic ranker is stage two.
  • Microsoft calls that stage L2 ranking, which “uses the context or semantic meaning of a query to compute a new relevance score over preranked results.”
  • It only reranks the shortlist, never the whole corpus: “What semantic ranker can’t do is rerun the query over the entire corpus… Semantic ranking reranks the existing result set, consisting of the top 50 results as scored by the default ranking algorithm.” Even if more than 50 results come back, “only the top 50 results progress to semantic ranking.”

Bing’s own May 2026 blog on the evolving role of the index doesn’t name reranking directly, but reinforces that retrieval quality is now judged by answer-support reliability: “Retrieval systems must therefore optimize not just for one-shot retrieval, but for consistent, repeatable behavior across iterative use.”

This is where reranking touches AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index., AI Mode, Copilot, ChatGPT Search, and Perplexity most directly. In a RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. pipeline, reranking is a named stage between retrieval and generation: content is chunkedChunking is splitting a document into smaller passages so AI systems can embed, index, and retrieve the single most relevant piece — not the whole page — in response to a query. It's a foundational step in RAG pipelines and the conceptual cousin of Google's passage ranking., each chunk is embedded and stored, the query retrieves nearby chunks by vector similarity, a reranker re-scores those candidates, and only the top survivors get handed to the LLM as context. The reranker is the gate between “your passage was retrieved” and “your passage was actually used.”

That gate can be strict. In AI-search systems, only a fraction of retrieved sources typically clear the rerank threshold into the generation stage — so being pulled into the candidate pool is the price of entry, not a guarantee of a citation. As Ahrefs’ own research on optimizing for LLM search frames the core problem: “AI companies don’t reveal how LLMs select sources, so it’s hard to know how to influence their outputs.” Reranking is a big part of that hidden selection step.

Reranking vs. Reciprocal Rank Fusion (RRF)

These get conflated constantly — including in otherwise-good SEO content — and they’re not the same mechanism.

  • Reranking rescores one candidate pool by jointly evaluating each query-document pair with a single model (the cross-encoder). It asks: how relevant is this document to this query, really?
  • Reciprocal Rank Fusion (RRF) merges multiple already-ranked lists — for example, the results from BM25 and the results from vector search, or the results from several fan-out sub-queries — by rewarding documents that appear consistently across lists. Ahrefs’ Query Fan-Out explainer describes it: fan-out queries are searched across indexes “using reciprocal rank fusion (RRF) — a method that scores and merges multiple lists of results by rewarding those that appear consistently across them.”

Both can live in the same pipeline — Azure’s semantic ranker literally reranks on top of a BM25- or RRF-ranked set — but RRF is a list-merging step (no model reads your content), while reranking is a content-scoring step (a model reads the query and your passage together). If you take one disambiguation away: RRF combines lists; reranking re-reads content.

A brief history: BM25 → RankBrain → BERT → LLM rerankers

Reranking isn’t new — it’s the modern name for a pattern search has used for years. The throughline, which I walk through in my Ahrefs Evolve 2025 talk GEO? AEO? LLMO? What’s With All This AI Stuff?:

  • BM25 / lexical retrieval — the classic keyword-match scoring that still does first-pass narrowing.
  • RankBrain (2016) — Google’s first machine-learning ranking system, understanding words as concepts.
  • BERT / DeepRank (2019) — contextual, passage-level language understanding; the cross-encoder-style reranking era begins.
  • Modern LLM-based rerankers (RankEmbed and RAG-era cross-encoders) — neural rerankers now sit between retrieval and generation across AI search.

The consistent shape across all of them: cheap broad retrieval first, expensive precise reordering of a shortlist second.

What this means for content and SEO

Because a cross-encoder scores the query and your passage jointly, the practical implications reinforce best practices you already know — now with a mechanism behind them:

  • Write self-contained passages. A reranker scores a passage largely on its own merits against the query. A section that only makes sense in the context of the three paragraphs above it scores worse than one that reads as a complete answer. This ties directly to passage rankingPassage ranking is a Google AI system that scores individual sections ('passages') of a page so a single page can earn multiple relevance scores for different queries. Google still indexes whole pages — only the ranking changed. and chunkingChunking is splitting a document into smaller passages so AI systems can embed, index, and retrieve the single most relevant piece — not the whole page — in response to a query. It's a foundational step in RAG pipelines and the conceptual cousin of Google's passage ranking..
  • Answer the specific question, near the top of the section. Direct answers score better than build-up. Put the answer first, then elaborate.
  • Minimize ambiguity. Pronouns and context-dependent phrasing (“as mentioned above,” “this approach”) that only resolve elsewhere on the page make a passage harder to score in isolation. Name the thing.
  • Retrieval is still the prerequisite. Reranking only ever sees what retrieval hands it. A page that can’t be crawledCrawling is how search engines use automated bots (like Googlebot and Bingbot) to discover URLs and download pages. A page has to be crawlable to be indexed, but crawling on its own isn't a ranking factor. and indexed, or that never gets retrieved, never reaches the reranker at all. Fix findability first; optimize passages second.

None of this is a knob you submit to Google. It’s the same “be clear and be found” advice, aimed at the specific stage — the second look — that decides which retrieved content actually gets used.

Add an expert note

Pin an expert quote

New person? Create their unclaimed profile at /admin/experts/ → Pin a quote first.