Chunking

When a machine reads your website, it does not read your website. It breaks each page into short passages — a few sentences, a heading and the paragraph beneath it, a list, a row of a table — and stores each one separately. Later, when someone asks a question, the machine searches those passages, not those pages. What comes back is a fragment, judged entirely on its own.

This is not a shortcut or a flaw. It is how retrieval has to work: a system answering a question needs the specific paragraph that answers it, not forty pages to wade through. The passage that gets pulled is the one that looked most relevant on its own terms, and it arrives at the answer without the page around it — no headline above it, no paragraph before it, none of the context you assumed a reader would carry down the page.

This is where ordinary, well-written pages quietly fail. A paragraph opening "as mentioned above, this is included in the price" says nothing once lifted out — what is included, and in whose price? A pricing table three sections below the heading that names the service has lost the service. A sentence beginning "we also do this for commercial clients" has lost the "this". Every pronoun, every backward reference, every idea split across a paragraph break is a place where the fragment arrives incomplete and gets passed over for one that did not.

Writing for extraction is mostly a matter of repeating yourself where you were taught not to. Name the subject again instead of saying "it". Put the answer directly under the heading that asks the question, not three paragraphs later. Keep one idea to a passage, and let each passage stand up alone. Give a table a caption that names what it contains. None of this makes a page worse to read; it mostly makes it clearer for people too.

The plain version: imagine any single paragraph of your site read aloud by a stranger, with nothing either side of it. Would a listener know what was being described, who was describing it, and whose business it was? That is the test the machine applies to every passage you publish, and it applies it thousands of times without ever once seeing the page whole.