What Is a Reranker, and Why Does It Decide Whether Your Page Gets Quoted?

Why does a page that ranks well still never get quoted by an AI answer?

Because ranking and retrieval are not the same contest. Most AI retrieval systems run in two passes. The first pass gathers a wide set of candidates cheaply. The second pass, called reranking, re-scores those candidates carefully and decides which few actually get used. Plenty of pages survive pass one and die in pass two.

This is the part of answer engine work that almost nobody writing about AI visibility explains, and it is the part that changed how I structure a page. If you understand what a reranker is doing, several pieces of common advice stop making sense and a few unglamorous habits start paying off.

What follows is what rerankers actually are, according to the companies that build and sell them, and what I think it means for anyone trying to get cited.

What is a reranker in plain language?

A reranker is a model that takes a question and a pile of candidate passages, then scores how relevant each passage is to that specific question. Voyage AI's documentation puts it directly: a reranker, given a query and many documents, returns the ranks of relevancy between the query and those documents.

The key word is relevancy to a query. A reranker does not score how good your page is in general. It scores how well one specific passage answers one specific question, at the moment that question is asked. The same passage can score highly for one phrasing and poorly for a near neighbour.

Cohere describes its own Rerank API endpoint the same way. Given a query and a list of documents, Rerank indexes the documents from most to least semantically relevant to the query, and Cohere positions the endpoint as a tool for semantic search rather than as a general quality model.

How is reranking different from the embedding search that comes first?

Embedding search compares two things that were turned into numbers separately. Reranking looks at the question and the passage together. Voyage AI's docs are explicit about this: rerankers are cross-encoders that jointly process a pair of query and document, unlike embedding models that encode queries and documents separately.

That difference sounds academic and is not. Encoding separately is fast and scales to millions of documents, but it compresses each passage into a single point before it has any idea what will be asked. Processing the pair together is slower and much more precise, because the model can attend to the exact words of the question while reading your paragraph.

This is why the architecture exists at all. You cannot afford the accurate method across an entire index, so you use the cheap method to get to a shortlist and the expensive method to order the shortlist. Voyage notes that applying a reranker to the top candidates from embedding-based search, or from lexical algorithms such as BM25 and TF-IDF, is common practice.

Why does the second pass change which page wins?

Because the two passes reward different things. Pass one rewards topical proximity, so a page broadly about your subject gets through. Pass two rewards a passage that directly addresses the question asked, so a page that is broadly about the subject and specifically about nothing loses to a narrower competitor.

I see this constantly when a comprehensive guide loses to a short, unglamorous page. The guide covers the question somewhere in its nine thousand words. The short page is the question. Under a scoring model that reads the query and the passage together, the short page wins on the only axis being measured.

The second reason is boundaries. A reranker scores whatever unit it was handed. If the unit it was handed starts mid-argument and ends before your conclusion, it scores a fragment of your thinking rather than your thinking. Nothing about the rest of your excellent page is visible to it.

Do ChatGPT, Perplexity and Google actually use rerankers?

Nobody outside those companies knows, because none of them publish their retrieval pipelines. What is documented is that reranking is a standard, commercially available stage that Cohere and Voyage AI both sell for exactly this purpose. Treat it as the shape of the problem, not as a claim about any specific product.

I want to be careful here, because a lot of AI visibility advice is confident about internals nobody has disclosed. The honest position is that two-stage retrieval is the dominant public architecture for search over large document sets, and that writing for it costs you nothing even if a given engine does something different.

What makes the bet safe is that everything reranking rewards is also what a careful human reader rewards. A self-contained paragraph that answers the question it raises is better for people regardless of what any pipeline does with it.

What do the reranker vendors publish about their own models?

Voyage AI lists rerank-3 and rerank-3-lite in preview with a 32,000 token context length, describing rerank-3 as its highest accuracy option and rerank-3-lite as optimised for latency-sensitive applications. It also lists rerank-2.5 and rerank-2.5-lite as generalist models with instruction-following and multilingual support, and a legacy rerank-2 with a 16,000 token context.

The context length is the number worth noticing. A 32,000 token window means the model can consider a substantial passage, not just a sentence. That cuts against the advice to chop everything into tiny fragments, because the scoring model has room to read a real section if a real section is what you gave it.

The split between a quality model and a lite model also tells you something about the economics. Reranking is the expensive stage, so systems apply it to a shortlist rather than to everything. Getting onto the shortlist and then winning the shortlist are genuinely two different problems, and most content advice only addresses the first.

What does this change about how you write a page?

Write sections that survive being read alone. Every heading should be a question a person would actually type, and the paragraph directly under it should answer that question completely, without depending on anything above it. That single habit does more for retrieval than any amount of keyword work.

The practical test I apply is to copy one section out of the page, paste it somewhere with no surrounding context, and read it as a stranger. If it needs the previous section to make sense, it will score badly when it is handed to a scoring model on its own. I keep rewriting until each section passes that test.

The second habit is to answer before elaborating. Context first, answer later is how most people write naturally and it is the wrong order here. I argued the specific version of this in how to write answer blocks that get cited, and understanding reranking is the reason the technique works rather than just a stylistic preference.

Why do chunk boundaries matter more than overall page quality?

Because the reranker never sees your page. It sees a slice. Where that slice begins and ends is decided by a chunking step you do not control, and the best you can do is make your natural boundaries and the likely slice boundaries line up as often as possible.

Headings are the main lever. A clear heading followed by a complete answer creates an obvious seam, and most chunking approaches respect obvious seams. Long undifferentiated prose creates no seams at all, so the slicing falls wherever the token count runs out, which is usually mid-thought.

I went deeper on the mechanics of this in content chunking for AI retrieval. The short version is that structure is not decoration here. It is the only influence you have over how your work gets cut up before it is judged.

What did I stop doing once I understood this?

I stopped writing transitional paragraphs that exist only to connect two sections, and I stopped burying the actual answer three paragraphs into a section for dramatic effect. Both were habits from writing for humans reading top to bottom, and both actively hurt a passage that gets read in isolation.

I also stopped treating word count as a proxy for thoroughness. A longer page gives you more chances to be retrieved and more chances to be diluted. What I care about now is how many sections in a piece could stand alone as a good answer, which is a completely different target from total length.

The last thing I dropped was the assumption that one great page could own a topic. Under two-stage retrieval, different phrasings of the same question can surface different passages, so coverage across several precise pages often beats depth in one sprawling one. That also changed how I think about which pages get pulled into an AI answer in the first place.

What should you do next?

Take your most important page, split it at the headings, and read each section as if it arrived with no page around it. Rewrite every section that cannot stand alone. Do that before you touch schema, internal links, or anything else on your AI visibility list.

If you want a second thing to do, look at whether your headings are questions someone would type or labels someone would skim. Labels are for navigation. Questions are for retrieval, and you can have both by writing the question and letting the design handle the skimming.

If you are trying to work out why a page you are proud of never gets quoted, send me the URL and the question you expect it to answer. I will tell you where I think it is losing. Let's chat.