Passage Ranking and Answer Blocks: Structuring Content for AI Extraction

Open a page that ranks for a real question and watch where Google’s AI Overview, Perplexity, or ChatGPT search actually quotes from. It is almost never the whole page and almost never the introduction’s throat-clearing. It is one tight block of text — usually 40 to 75 words — that states a complete answer without requiring anything around it. That block is the unit of AI visibility now, and most content is not written in blocks at all. It is written as a continuous argument that assumes the reader starts at the top and proceeds in order. Models do not read that way. They retrieve passages.

Passage Ranking Changed the Unit of Retrieval

Google announced passage ranking in 2020 and rolled it into production in 2021. The mechanism is straightforward and consequential: the ranking system can identify and rank an individual passage from within a page, independent of how well the page as a whole targets the query. Before passage ranking, a comprehensive 4,000-word guide could lose to a thin dedicated page because the relevant section was buried among unrelated content and the page-level signals were diluted. After passage ranking, that buried section can surface on its own merits.

AI answer engines generalize this further. A retrieval-augmented system — which is what ChatGPT search, Perplexity, and Google’s AI Overviews all are underneath — does not load your entire page into the model’s context and ask it to reason over the whole thing. It chunks your content, embeds each chunk as a vector, retrieves the chunks most semantically similar to the query, and grounds the generated answer on those chunks. The chunk is the atom. If your answer is split across three paragraphs separated by tangents, the retriever may pull only one of them, and the model generates from an incomplete fragment — or skips your page for one where the answer sits in a single coherent chunk.

This is the structural fact that should drive how you write for AI extraction: the model rewards content that is already decomposed into self-contained, retrievable units. You are not optimizing a page. You are optimizing the individual passages within it.

The Self-Contained Answer Block

The single highest-leverage move is to place a complete, self-contained answer immediately after the heading that poses the question, near the top of the page. “Self-contained” is the operative term and the one most writers get wrong. A self-contained answer resolves the question using only the words in that block — no antecedents that live three paragraphs up, no “as discussed above,” no pronoun whose referent is in a different section.

Compare two openings to a page titled “How long does a 301 redirect take to pass link equity?”

The diluted version: “Redirects are a critical part of any site migration. There are several types, and choosing the right one matters. In this guide we’ll cover everything you need to know about how they work and what to expect.” A model retrieving this chunk learns nothing about the actual answer. It will not cite it, because there is nothing to cite.

The self-contained version: “A 301 redirect passes the large majority of link equity almost immediately once Googlebot recrawls the source URL, typically within days to a few weeks depending on crawl frequency. The redirect should stay in place for at least a year — Google has stated that link signals consolidate fully only after the target has been crawled and indexed in place of the source over multiple crawl cycles.” That block answers the question, includes the qualifier, names the mechanism, and stands alone. It is liftable. A model can ground an answer on it verbatim and attribute it to you.

The length target — roughly 40 to 75 words — is not arbitrary. Below about 30 words you rarely have room for the claim plus its necessary qualifier, and an unqualified claim reads as either obvious or wrong. Above about 90 words the block stops being a clean extractable unit and becomes a paragraph the model must summarize, which reintroduces paraphrase and reduces the odds your exact phrasing surfaces. The sweet spot is one to four sentences that fully resolve one question.

Why the Lead Beats the Wall of Text

The instinct to “build up” to an answer — context first, nuance, then the conclusion — is the enemy of extraction. It is also, separately, bad for human readers, which is why news writing abandoned it a century ago in favor of the inverted pyramid: conclusion first, supporting detail after, least important material last.

A wall of text forces the retriever to choose a chunk boundary inside your reasoning. Chunkers split on token counts or paragraph breaks, not on semantic completeness. If your answer emerges only in the synthesis of paragraphs four through six, the retriever has no single chunk that contains it. It pulls paragraph five, which contains a mid-argument fragment, and either grounds a weak answer on it or moves on. The lead-first structure guarantees that the most retrievable position on the page — the top — contains the most complete answer. You are placing the payload where the retriever looks first and where the chunk boundary is cleanest.

This does not mean abandon depth. It means front-load the answer and let the body earn the trust that makes the answer credible. The lead states the conclusion; the sections beneath it supply the evidence, the edge cases, and the mechanism. Readers who want only the answer leave satisfied; readers who want the reasoning stay. Both audiences — and the model — are served by the same structure.

Heading Hierarchy Is the Retrieval Index

Headings are not decoration and they are not only for humans skimming. For both Google’s passage systems and AI retrievers, the heading is the strongest local signal of what the passage beneath it is about. A well-formed heading hierarchy is, functionally, a table of contents the retriever uses to understand which passage answers which question.

Three rules govern heading structure for extraction. First, exactly one H1, and it should name the page’s primary entity or question — not a brand slogan, not a clever phrase that omits the topic. Second, H2s should map to the distinct questions or sub-topics a reader (or a model) would ask about the H1 topic, phrased the way the question is actually asked. A heading reading “How long does it take?” is weaker than “How long does a 301 redirect take to pass link equity?” because the latter is self-contained at the heading level — it carries its own context, exactly as the answer block beneath it should. Third, do not skip levels. An H4 under an H2 with no intervening H3 breaks the implied hierarchy and confuses both the accessibility tree and the structural parse a retriever builds.

The payoff: when each H2 poses a complete question and the 40-to-75-word block beneath it answers that question self-containedly, every section of your page becomes an independently retrievable answer unit. A 2,000-word page structured this way is not one document competing for one query. It is a dozen answer blocks, each competing for its own query, each liftable on its own.

Answer Blocks Are Not FAQ Schema

A common misread is to assume this means stuffing the page with FAQPage structured data. It does not. Google deprecated FAQ rich results for most sites in 2023, and AI engines extract from your visible body content, not from a JSON-LD block they may or may not parse. The answer-block discipline is about the prose a human and a model both read — the H2 question and the self-contained paragraph under it — not about marking it up as FAQ schema. Schema helps machines understand entities and relationships; it does not make a buried answer retrievable. The retrievable answer has to exist in the visible text.

This is also why raw-HTML availability is the prerequisite beneath all of this. Most AI crawlers do not execute JavaScript. If your beautifully structured answer block is injected into the DOM after hydration, the retriever fetches the page and finds a shell where the answer should be. The block has to be present in the server-rendered HTML, in the response the crawler actually receives, or none of the structure matters. Render parity comes first; passage structure comes second; extraction comes third. Skip the first and the rest is invisible.

The Practical Pattern

Take any page you want cited. For its primary question, write a 40-to-75-word self-contained answer and place it directly under the H1. For every distinct sub-question, create an H2 phrased as the full question and put a self-contained answer block immediately beneath it. Read each block in isolation — literally cover the rest of the page — and confirm it answers its heading without reaching elsewhere. Remove pronouns whose referents live in other sections. Confirm every block exists in the raw HTML, not just the rendered DOM.

Then check the boundaries. If two adjacent paragraphs each contain half an answer, merge them into one block or split the question into two headed sections, each with its own complete answer. The goal is that no answer ever straddles a chunk boundary, because you cannot control where the retriever draws that boundary — you can only control whether a complete answer sits cleanly within one.

A visibility platform helps here by analyzing whether your pages are structured as retrievable answer units rather than continuous prose: whether each heading poses a self-contained question, whether a complete answer block follows it, whether those blocks survive in the raw HTML a non-rendering AI crawler receives, and whether heading hierarchy is well-formed enough to serve as a retrieval index. It does not promise a citation — no honest tool can, because retrieval and grounding decisions belong to engines you do not control. What it can do is tell you whether your content is structured to be extractable and citable in the first place, which is the part you can actually fix.