How chunking works
A retrieval system almost never stores a whole page as one item. It cuts the text into passages, turns each passage into an embedding, and stores them separately. When a question arrives, the system compares the question with those passages and returns the closest few — so what competes for a place in an answer is a chunk, not a page.
Where the cuts fall is the whole game. Naive splitting counts characters and slices wherever the count runs out, which cheerfully cuts a sentence in half. Better splitting follows the document’s own structure — headings, sections, list items, table rows — so each piece is about one thing. Many systems overlap adjacent chunks slightly so a sentence at a boundary appears in both, and most attach the source URL and title to every piece so a retrieved passage can still be credited.
Why chunking matters
It explains something publishers find counter-intuitive: a strong page can lose to a weaker one because the weaker one is easier to cut cleanly. If your answer only makes sense alongside the paragraphs above it, the chunk containing it reads as fragmentary and gets passed over. If the same answer is stated once, completely, under a heading that names the question, the chunk stands up on its own.
The same logic applies inside your own systems. If you feed a chatbot your PDFs and price lists, chunking is what decides whether it answers a customer’s question correctly or returns half a table. The quality of the source document matters far less than the tidiness of its structure.
Where chunking goes wrong
The usual damage comes from pronouns. A passage that opens with “it also includes” or “this is available in” has lost its subject the moment it is separated from the section above, and a model reading it in isolation may attach the claim to the wrong service. Tables split across a boundary do something worse: the header row goes one way and the values the other, and figures end up attributed to nothing.
The other frequent mistake is the everything page — a single long service page covering audience, process, pricing model and frequently asked questions. It may read well to a human scrolling, yet every chunk cut from it is a mixture, and mixtures match nothing precisely.
How to act on it
Write so that any section could be photocopied and still make sense. In practice that means a few small habits:
- Give each question its own heading, phrased as the question.
- Answer in the first sentence under that heading, before context.
- Repeat the subject noun instead of writing “it” or “this”.
- Keep a table and its explanation together, and keep the table small.
Split genuinely separate topics onto separate pages rather than stacking them under one URL, and let the length follow the topic. None of this is a trick for machines — it is the same structure that helps a reader skim, which is why it also improves retrievability without any technical work at all.