What is chunking and why does it happen?
Chunking is what retrieval systems do to a page before they can search it: they split the text into pieces of a manageable size, often along heading boundaries, and index each piece separately, typically as an embedding. When a query arrives, the system finds the chunks most similar to it, not the pages, and passes those chunks to the model that writes the answer.
It happens because models have limited context and pages are long. The consequence for a writer is that the page as a whole is never what the engine judges; the section is.
What makes a chunk lose?
A chunk loses when it does not make sense on its own. A section that begins with as we saw above, or refers to it and this without naming the entity, or contains half of a comparison whose other half is in the next section, cannot answer a query by itself and is ranked below one that can.
A chunk also loses when its answer is buried. If the section's first two sentences are preamble and the answer is in the fourth, a competing chunk that opens with the answer looks more relevant to the query and is chosen.
What are the rules for chunk-safe writing?
Restate the entity and the topic in each section's first sentence, so the chunk names what it is about. Put the answer first, then the reasoning, then the specifics. Avoid every cross-reference: no above, below, earlier, or later. Keep each section to roughly 120 to 220 words, which is long enough to answer and short enough to be a single chunk rather than split mid-argument.
Give each section a heading that is a question or a claim, because headings usually travel with the chunk and tell the retriever what the chunk answers. A heading like benefits tells it nothing; how does the approval gate work tells it everything.
- First sentence restates entity and topic
- Answer first, reasoning second, specifics third
- No cross-references to other parts of the page
- 120 to 220 words per section
- Heading is a question or a claim
How do you test a section?
Test a section by reading it with everything else on the page hidden. Ask three questions: what entity is this about, what question does it answer, and could a reader act on it without the rest of the page. If any answer is unclear, the chunk is not self-contained.
A second test is to imagine the section quoted alone in an AI answer, attributed to your site. If it would be confusing or misleading out of context, rewrite it until it would not. That is precisely how it will appear when cited.
How does this fit the rest of the discipline?
Chunk-safe writing is the mechanical basis of the rest of AI-search discipline. Answer-first openings, question headings, entity restatement, and fan-out coverage are all ways of making chunks that win. The SPEC-035 standard on ftes.ai states it directly: every section is a standalone chunk, because retrievers rank sections, not pages.
It also produces better writing for humans. A section that makes sense on its own is a section a skimming reader can land on and understand, which is how most readers actually read.
How does chunk discipline change the way a piece is drafted?
Chunk discipline changes drafting from writing a page to writing a set of sections that happen to sit on one page. The Content FTE on ftes.ai starts from the brief's list of sub-queries, writes each section to answer its own heading with the entity named in the first sentence, and only afterwards checks that the sections read well in sequence. Transitions are allowed to be light because no section may depend on them; a reader landing on section four from a search result must not need section three.
The fact gate then checks claims per section, because a claim with its source in a different section is unsourced in the chunk an engine retrieves. Editing in Approvals follows the same rule: a send-back that says a section only makes sense after the previous one is a structural fault, not a style note, and the fix is to restate the entity and the context inside the section. Drafted this way, a piece is a cluster in miniature: every section is a candidate for a citation on its own, and the page is the sum of them rather than a container for them.