Before a retrieval-augmented system can use your page, it splits it. Understanding how, and writing to survive the split, is concrete work with a clear mechanism — unlike most advice about "optimizing for AI."
How splitting works
Most production pipelines use one of a few strategies:
Fixed-size chunking. Split every N tokens, often with overlap between adjacent chunks. Simple, fast, and indifferent to your document structure — it will cut mid-sentence and mid-argument.
Recursive character splitting. Try to split on paragraph breaks; if a chunk is still too large, split on sentences; then on words. Respects structure when it can. This is the common default in widely used frameworks.
Semantic chunking. Embed sentences, detect topic boundaries by similarity drop, split there. Better quality, more expensive, less common in production.
Structure-aware chunking. Split on document structure — headings, sections, list boundaries. Best results for well-structured HTML and Markdown, and increasingly common for web content specifically because HTML carries that structure explicitly.
The important consequence is shared across all of them: the unit of retrieval is a passage of a few hundred words, not your page. Your carefully constructed argument spanning six paragraphs may be retrieved as paragraph four, alone, with no surrounding context.
Writing that survives it
Make sections self-contained. Each section should establish its own subject. "This means the second option is usually better" is unusable alone. "Zirconia crowns generally last longer than porcelain-fused-to-metal crowns, though they are harder to adjust after placement" survives extraction intact.
The practical test: read any single section as if it were the first thing you encountered. If it depends on knowing what came before, rewrite the opening sentence to restate the subject.
Front-load the answer. Put the conclusion in the first sentence or two of a section, then support it. If a chunk boundary falls mid-section, the half containing the answer is still useful.
Use headings as retrieval anchors. Structure-aware chunkers attach the heading to the chunk. A heading that restates the question in the reader's language — "How long does a dental implant take from start to finish?" — makes the chunk match that query far better than "Timeline" does.
Keep sections tightly scoped. A section covering three topics produces a diffuse embedding that matches none of them well. One idea per section produces a clean vector.
Repeat key nouns instead of pronouns across section boundaries. Within a paragraph, pronouns are fine. At the start of a new section, "it" and "this" refer to something the chunk may not contain. Name the thing.
Put facts in structured elements. Tables and lists survive chunking unusually well because they are self-labelling. A table of costs by material carries its own column headers into whatever chunk it lands in.
What this looks like in practice
Weak structure for retrieval:
## Overview There are several factors to consider... ## More on this As mentioned, the second point is particularly relevant because...
Strong structure:
## What determines the cost of a dental crown Material is the largest single factor. Porcelain-fused-to-metal crowns typically cost less than all-ceramic or zirconia crowns because... ## Does dental insurance cover crowns Most dental plans cover roughly half the cost of a crown when it is medically necessary, subject to the plan's annual maximum...
The second version is also better for human readers, better for featured snippets, and better for anyone skimming. That alignment is not coincidental — all of these consumers benefit from the same property, which is that each section answers one question completely.
What does not help
Keyword stuffing inside chunks. Embedding models capture meaning, not term frequency. Repetition changes the statistics without changing the meaning.
Artificially short sections. Splitting one coherent idea into three headed fragments produces three weak chunks instead of one strong one. Match section boundaries to idea boundaries, not to a target length.
Adding a summary block at the top and assuming it will be retrieved. It might be. So might section four. Write every section as if it is the one that gets pulled, because you do not control the choice.
The honest scope of this advice
There is no confirmed ranking mechanism here, and no vendor has published how their retrieval weights passages. What is documented is how chunking works, and what is observable is which passages get quoted in AI answers.
The recommendation is defensible on that basis: structure content so that any section can be understood alone, because that is demonstrably the unit these systems consume. It also happens to produce better-written pages, which means the downside if the mechanism shifts is approximately zero.