Skip to content
Citable

Definition

Passage retrieval

Passage retrieval is the search technique of scoring and returning short spans of text — a paragraph, a section, a window of tokens — rather than whole documents, so that a system can answer a question from the most relevant fragment. It is the retrieval step in most answer engines, and it is why a page is cited passage by passage rather than as a unit.

Passage retrieval comes from open-domain question answering, where the task is to find the few sentences in a large corpus that answer a question. Dense Passage Retrieval (Karpukhin et al., 2020) showed that embedding passages and questions into one vector space could beat the classic BM25 keyword baseline; that method, usually combined with keyword matching and a reranker, sits inside most retrieval-augmented generation pipelines today. Google announced passage-level ranking for ordinary web search at its Search On event in October 2020.

The practical consequence is that your page is never judged whole. It is cut into chunks, each is scored against the question, and only the winners reach the model. The heading above a paragraph, its opening sentence and whether it names its subject explicitly decide whether it is a candidate at all.

Making a passage retrievable

  • One question per section, with the question or its key phrase in the heading
  • The answer in the first sentence, then the qualification
  • The full entity name in the passage, not a pronoun pointing at an earlier section
  • Figures, units and dates inside the same paragraph as the claim they support
  • Plain HTML paragraphs and lists, so the chunk boundaries fall where the meaning does

Modelling content so that every chunk is self-contained is covered in Modelling content for answer engines in a headless CMS.

Frequently asked questions

Why does passage retrieval favour answer-first writing?

Because the retriever scores each passage on its own, without the rest of the page. A paragraph that opens with the answer, names the entity in full and needs no earlier context resembles the question closely and is complete when lifted out. A paragraph that builds up to the answer scores lower and, if selected, gives the model a fragment that stops mid-thought.

Related terms

Guides that use this term

  • Content

    How to write answer-first content that LLMs can quote

    Answer-first content opens every page and every section with a self-contained answer of roughly 40–70 words, then explains, qualifies and expands. The first sentence names the main entity, each heading carries exactly one question, and no passage depends on a pronoun or a paragraph elsewhere. Written this way, a passage still makes sense after an answer engine splits the page into chunks and retrieves one of them in isolation.

    8 min read

  • Fundamentals

    What is Answer Engine Optimization (AEO)?

    Answer Engine Optimization (AEO) is the practice of structuring content so that AI-powered answer engines — ChatGPT, Perplexity, Claude, Copilot and Google's AI Overviews — can retrieve it, understand it and cite it inside a generated answer. Where classic SEO competes for a ranking position, AEO competes for extraction: a self-contained passage that resolves one question, from a page with a clear author, date and sources.

    6 min read

  • Technical

    Modelling content for answer engines in a headless CMS!

    A headless CMS content model for answer engines makes every Answer Engine Optimization signal a constrained field, not an editorial habit: a required short answer with a maximum length, key takeaways, FAQ items, sources with URL validation, the author as a relation, and published and updated dates. JSON-LD, Markdown and llms.txt are then generated from those fields, so no representation can disagree with the page.

    9 min read