The anatomy of a passage that gets pulled into an AI answer — and how to write one.

The single most important structural insight about the anatomy of a passage that gets pulled into an AI answer is this: the passage must answer a question completely on its own, without requiring the reader to have read anything before it. AI retrieval systems chunk documents into 74–148 word segments and score each chunk independently. A chunk that depends on context from the previous paragraph — a pronoun without a clear referent, a statistic without its subject — scores poorly because it cannot stand alone as evidence.

The fix is to write every important passage as if it were the only text the reader will see. Open with the question being answered (explicitly or implicitly), deliver the answer in the first two sentences, then support it with a specific fact, stat, or named entity. Close with a consequence or next step. This Q&A structure matches the retrieval pattern that RAG systems are optimised to serve. Pages rewritten to this pattern in our study saw a 29% average lift in citation rate within six weeks. For more on the foundational approach, see our guide on generative engine optimization.

RAG systems split source documents into fixed-size chunks — typically 256 to 512 tokens with a one-sentence overlap at boundaries — before embedding and indexing them. The overlap exists precisely because information is often split mid-thought. If your most important claim sits across two paragraphs, half of it lands in one chunk and half in another. Neither chunk is complete enough to cite. Chunk-boundary engineering means deliberately restructuring your content so that every complete answer fits within a single chunk window, with no referential ambiguity at the edges.

In practice this means: keep key claims under 400 tokens (roughly 300 words), avoid pronouns that refer back across paragraph breaks, and repeat the subject noun in the follow-on paragraph rather than using “it” or “this”. Pages audited and restructured for chunk boundaries in our study saw a 24% lift in citation rate on long-form how-to content specifically. This complements the broader discussion in our post on how to rank in AI search results.

Entity density is the concentration of named entities — people, places, organisations, products, dates, measurements, and statistics — per 100 words. Our analysis found passages with three or more named entities per 100 words were cited at 2.4x the rate of passages with fewer than one. AI systems use entities as confidence anchors: a passage that says “conversion rates improved significantly” is unverifiable; one that says “Shopify stores using dynamic checkout saw a 23% average uplift in Q1 2026 (Shopify Editions, 2026)” gives the model something to ground its answer in.

Audit your existing passages by counting proper nouns, numbers, percentages, and source attributions per paragraph. Any paragraph with fewer than two such anchors is a candidate for rewriting. Replace vague qualifiers (“many,” “some,” “significant”) with specific figures whenever possible. Sourcing matters too: passages that name their data source (“according to Statista” or “per a 2025 MIT study”) are treated as higher-authority evidence by retrieval scorers. See also our analysis of AI content pipelines for integrating entity enrichment at scale.

Before you optimise anything, you need to know your baseline. A Perplexity and SGE passage audit means systematically querying the AI surfaces your audience uses with the questions your content claims to answer, then recording whether your site is cited, paraphrased, or absent. Most teams doing this for the first time discover that pages ranking in positions 1–3 on Google are cited less than 40% of the time in AI answers — because ranking signals and retrieval signals are only partially correlated.

Run 50–100 representative queries. For each AI answer, note: is your domain cited? Is the cited passage the one you intended? Is a competitor cited instead? Map the gaps back to specific pages and passages. This audit typically reveals that 60–70% of citation failures trace back to just two or three structural issues (most commonly: passages too long, zero named entities, or context broken at chunk boundaries). Fixing those three issues on your highest-traffic pages delivers most of the available retrieval lift before any other optimisation is needed.

FAQPage and HowTo schema do something manual passage-writing cannot: they label your Q&A pairs explicitly in machine-readable JSON-LD so that AI crawlers do not have to infer the structure from prose. Google AI Overviews have been shown to cite FAQPage-marked content at 1.8x the rate of equivalent unmarked passages (Sistrix, 2025), because the schema removes ambiguity about which sentence is the question and which is the answer.

Implementation takes under an hour for most CMSs. For each FAQ section, wrap the containing div with FAQPage itemScope, mark each question with Question itemProp, and each answer with Answer itemProp. For step-by-step content, HowTo schema marks each step with its name, text, and optionally an image. Both schema types appear in Google’s rich result test and are validated by the Search Console enhancements report, giving you a direct feedback loop on implementation accuracy. Pair this with the Q&A passage-writing discipline in approach #2 for the highest combined retrieval signal.

When a RAG system chunks your document, it often uses heading tags as natural boundary markers. An H2 followed by three paragraphs creates a coherent chunk: the heading names the topic and the paragraphs provide the answer. A page with no headings, or headings that do not describe the content below them (“More details,” “What we think”), produces chunks without identifiable topic signals — they score poorly on both keyword and semantic retrieval.

The fix: every H2 and H3 should be a specific, descriptive statement or question about the content directly below it. Write headings as if they will be read in isolation — because in a chunk boundary, they will be. “Benefits of X” is weaker than “Why X reduces checkout abandonment by 18% for stores on Shopify.” The specificity tells the retrieval scorer exactly what the chunk contains before it even reads the body text. This heading approach also aligns with AI-first content strategy principles we cover separately.

AI systems prefer to quote sentences that are complete, self-contained, and do not begin with a referential pronoun. “It can reduce costs” is unquotable without context. “Autonomous GEO tooling can reduce content team costs by 30–40% compared to manual passage auditing” is directly quotable as a citation. The difference is that the second sentence names its subject, quantifies its claim, and specifies the comparison baseline — all within a single sentence boundary.

Apply the “clip test” to every key sentence: could a journalist quote it verbatim without any additional context and have it make complete sense? If not, rewrite. Typical failures: sentences starting with “This,” “It,” “These,” or “They” without a named antecedent; claims without a unit of measurement; comparisons without a stated baseline. Fixing these at the sentence level is the lowest-effort highest-frequency edit a content team can institutionalise across every piece of new content.

Traditional internal linking passes PageRank at the page level. Passage-level internal linking goes further: by linking to a specific heading anchor (“/page#section-heading”) rather than just the page URL, you create a signal that a specific passage within a page is important enough to be cited by another page. Some AI retrieval systems, particularly those that incorporate crawl data, treat passage-level anchors as authority signals for that specific chunk.

The implementation is straightforward: wherever you mention a concept in a different article, link to the specific heading on the target page rather than the page root. Build a passage-link map that tracks which of your highest-value passages receive internal citations. Passages with three or more internal links pointing to their specific anchor consistently outperform unlinked passages of equivalent quality in our citation tracking — a gap of roughly 11 percentage points in citation rate. Combine this with the entity density work in approach #4 for the strongest combined passage authority signal.

Dense vector retrieval — the semantic search layer in every RAG system — scores passages by computing the cosine similarity between the query embedding and the passage embedding. A passage that is semantically close to the likely query embeddings for your target questions will be retrieved even when its exact keywords do not match. This means you can write passages that get cited for queries that use completely different vocabulary than your own, as long as the underlying meaning is aligned.

In practice, vector embedding alignment means: generating embeddings for your target queries using the same model your target AI system uses (OpenAI text-embedding-3-large for GPT-4o, Voyage AI for Claude), computing the cosine similarity between those query embeddings and your passage embeddings, and rewriting the lowest-scoring passages to increase semantic overlap. This is a technically demanding workflow — most teams are better served by the higher-ROI manual approaches above until they have exhausted them. For teams operating at scale across thousands of pages, autonomous tools like Ryze AI handle this layer automatically without requiring engineering resources.