Content chunking is the step in which a retrieval system splits a page into smaller pieces before storing and searching them. When an AI assistant quotes your site, it is quoting one of these chunks. You cannot choose how a search engine chunks your page, but you can structure the page so that almost any sensible split produces pieces that make sense on their own.
How systems split pages
| Method | How it works | Effect on your page |
|---|---|---|
| Fixed size | Cuts every N tokens, often with some overlap between neighbouring chunks | Can cut a paragraph or table in half |
| Recursive splitting | Tries to cut at section breaks first, then paragraphs, then sentences, until pieces are small enough | Respects your structure when headings and paragraphs are clear |
| Structure-based | Uses headings (H2, H3) or HTML elements as boundaries | Each section under a heading becomes a chunk |
| Semantic | Compares the meaning of consecutive sentences and cuts where the topic shifts | Mixed-topic sections get split in unpredictable places |
Open-source retrieval frameworks such as LangChain and LlamaIndex ship default chunk sizes in the range of a few hundred to about a thousand tokens. Commercial search systems do not publish their settings. That is why advice to write in chunks of exactly 256 or 512 tokens is a rough guideline; no search engine has announced such a rule.
What makes a good chunk
- It covers one topic
- It starts by stating its main point
- It names its subject instead of relying on “it”, “this” or “the above”
- It includes the numbers, conditions or examples needed to make its claim true
- It is short enough to survive being cut to a few hundred words
Before and after
This section from a typical service page depends on the text before it:
As we covered earlier, this is where most clients start. It usually takes two weeks and the output is used by the team afterwards.
Retrieved on its own, a system cannot tell what “this” is or what the output contains. Rewritten as a self-contained chunk:
The AI Search Audit is the usual starting point for Brilad clients. It takes about two weeks and ends with a fix list ranked by impact, which the client’s own team can use.
Headings do double duty
A heading tells readers what a section covers and gives structure-based splitters a clean boundary. Headings phrased the way people ask (“How long does an audit take?”) also help retrieval match the chunk to a question. Vague headings such as “Overview” or “More details” do neither.
Tables and lists
Short tables usually stay in one chunk and are easy for models to read. Very long tables may be cut, so repeat the column headers if a table runs past thirty or forty rows, or split it into smaller tables with their own headings. Lists should introduce their topic in the sentence before them, so a chunk that begins with the list still has context.
Frequently asked questions
Should every page be split into very short sections?
No. Sections should be as long as the topic needs. What matters is one topic per section and a clear opening sentence.
Does chunking affect Google rankings?
Google has its own passage-based ranking, introduced in 2021, which can rank a single section of a long page. Clear, self-contained sections help there too.