No answer engine reads your page top to bottom. Retrieval pipelines cut documents into chunks — passages of a few hundred tokens, in most systems between 256 and 512 — embed each chunk as a vector, and match user questions against chunks, not pages. The page is a container; the chunk is the product. Writing that ignores this loses citations to writing that respects it, at identical content quality.
Headings are retrieval queries in disguise
Chunkers split on structure, and H2/H3 headings are the strongest split signal a page offers. After the split, the heading travels with its section text into the embedding — models such as multilingual-e5-small, a 384-dimension embedder covering 100 languages, encode the pair as one passage. A heading that promises one thing over a body that delivers another drags the whole chunk toward the wrong queries and away from the right ones. Descriptive, literal headings beat clever ones in retrieval every single time.
Each section lives or dies alone
A chunk arrives at the language model with no surrounding context: no previous paragraph, no earlier definition, no antecedent for that opening "this". Sections that open with "This approach also…" or "As mentioned above…" turn into orphans at retrieval time — grammatically fine, semantically empty. Open every section with its subject: name the thing, then discuss it. The habit costs nothing at writing time and decides whether the passage is quotable on its own.
Length matters for the same reason. A two-line section embeds into a vector that says almost nothing and matches almost nothing. Give every heading enough substance to answer at least one concrete question without help from its neighbours.
Redundancy splits your own vote
Two sections that restate the same point produce two nearly identical vectors. At query time they compete against each other for the same retrieval slots, splitting relevance mass that a single strong section would concentrate. Consolidating duplicated passages is one of the cheapest retrieval wins available: merge the sections, keep the best examples from each, and redirect the freed heading to a question the page does not answer yet.
Markup that helps the chunk travel
Structure survives chunking better than prose. Tables keep units attached to numbers, definition lists keep terms attached to meanings, and an Article object with real dates tells the engine which version of the facts it is quoting. Google’s structured data documentation frames markup as machine-confidence: the same claim, made twice, once for humans and once for parsers. Chunks inherit that confidence — and pages that provide it get quoted with fewer distortions.
Sources
The model card documents the embedder’s 384 dimensions and 100-language coverage; the schema.org and Google pages define the markup that keeps chunks attributable. All three links point at canonical, maintained documents.