Reranking

What Reranking Actually Is

Modern retrieval runs in two stages, and reranking is the second one. The first stage is fast and shallow. A dense retrieval system built on a bi-encoder like SBERT has already turned every passage into a vector in advance, so at query time it embeds the question once and grabs the nearest few dozen or few hundred passages by vector distance. This is quick because the question and the documents are encoded separately and never actually read side by side. That separation is exactly what makes it fast, and exactly what makes it imprecise.

A reranker fixes the imprecision. It is almost always a cross-encoder: instead of comparing two independent vectors, it feeds the query and a single candidate passage into the model together, in one forward pass, and lets every word of the question attend to every word of the passage. The output is one number, a relevance score for that specific query-passage pair. Run the cross-encoder over each candidate in the shortlist, sort by score, and you have a new, far more accurate ordering. Google's original BERT worked this way, and dedicated rerankers such as the Cohere Rerank models and open cross-encoders on Hugging Face are built for precisely this job.

A concrete example makes the difference clear. Ask "what are the side effects of switching accounting software mid-year," and a bi-encoder might rank a passage that merely mentions "accounting software" above one that directly discusses the risks of a mid-year switch, because in isolation both look topically close. The cross-encoder, reading the question and each passage jointly, sees that only the second passage actually answers the intent, and promotes it. The bi-encoder measured similarity of topic. The cross-encoder judged relevance to the question.

Why Two Stages Exist At All

The obvious question is why an engine does not just run the accurate cross-encoder over everything and skip the coarse first pass. The answer is cost. A cross-encoder has to process the query paired with each candidate individually, so scoring one million passages means one million full model passes at query time. That is impossibly slow for a live answer. A bi-encoder, by contrast, does all of its heavy work in advance and reduces query time to a single embedding plus a nearest-neighbor lookup, which is why retrieval-augmented generation is feasible over the open web in the first place.

So the two stages split the labor by their strengths. The bi-encoder is the fast filter that cheaply narrows millions of passages down to a manageable shortlist of, say, the top fifty. The cross-encoder is the expensive judge that reads that shortlist carefully and puts it in the right order. Speed first, then precision. It is the classic engineering compromise, and it is why almost every serious retrieval pipeline, from enterprise search to the systems behind AI answer engines, is a retrieve-then-rerank design rather than one model doing both jobs.

Why This Decides Whether AI Can Cite You

Here is the truth that matters most for visibility, and it is structural, not stylistic. Reranking operates only on the shortlist. A cross-encoder never sees the millions of passages the first pass discarded. It reads and reorders the fifty candidates the bi-encoder handed it, and nothing else. That means retrievability is upstream of everything. A page that is not retrieved in the first pass is not on the shortlist, is never reranked, is never handed to the language model, and is never cited, no matter how precisely relevant it might have been if only the reranker had been allowed to read it.

This reframes where the real competition happens. It is tempting to obsess over being the single best answer, the passage a perfect judge would rank first. But the reranker is not the gate. The first-pass retriever is. Your content must first embed near the questions your customers ask, well enough to survive the coarse cut, before its precise relevance can ever be rewarded. Reranking decides the order of the finalists. Retrieval decides who becomes a finalist. You cannot win the second contest if you never enter the first.

What This Means for Your Website

The practical work splits cleanly across the two stages, and both matter. To survive the first pass, your passages need to be retrievable: self-contained, answer-shaped, and embedded near real questions rather than buried in vague, sprawling prose. That is a job of embedding your meaning cleanly, so dense retrieval pulls you onto the shortlist in the first place. If you fail here, nothing downstream can save you.

To then win the rerank, the passage that gets retrieved must directly and unambiguously answer the specific question, not merely orbit the topic. A cross-encoder rewards passages that state the answer plainly, with the subject named and the claim complete inside the passage itself. Passages that hedge, that answer a slightly different question, or that lean on context three paragraphs away tend to score lower once a model reads them jointly with the query. Clarity is not decoration here. It is what the precision stage is measuring.

How AIOInsights Reads This Signal

AIOInsights does not grade you on the word "reranking." It grades the conditions that decide whether you clear the first pass and then hold up under the second. That work sits inside the AI Discoverability pillar, which evaluates whether your content is structured to be retrieved, shortlisted, and read: self-contained passages, explicit subjects, plain-text facts, and answer-shaped writing. It also draws on retrieval legibility, our term for whether your facts are stated in text a machine can actually read and score, because a fact locked in an image can be neither retrieved nor reranked.

Every one of those checks is real and deterministic. We do not run a live cross-encoder against your pages and report a number that drifts on each request. We evaluate the fixed, observable signals that decide, before any engine runs, whether your content survives retrieval and earns the reranker's attention. You can read the full method in our scoring methodology.

Check How Discoverable Your Site Is

Keep reading the lexicon: Dense Retrieval, RAG, and SBERT.