Retrieval-augmented generation

Open book instead of from memory

Picture two versions of the same test. In the first, a student answers from memory. In the second, the student is handed four relevant pages and told to answer out of them and say which page each part came from. The second student is more accurate, and you can check the work.

That is the difference retrieval makes. Left alone, a language model writes from patterns it absorbed in training. With retrieval bolted on, the system runs searches first, grabs a handful of pages that look relevant, and writes the answer from what those pages actually say.

So when a customer asks an assistant who repairs boat lifts near Cape Coral, two things happen in order. Something searches. Then something writes. Your business has to survive the first step to appear in the second.

The plumbing behind this keeps changing and the names for it vary by platform, but the two-step shape has held. It is also the reason AI answers cite sources at all. A model writing from memory has nothing to point at.

Retrieval reads passages, not websites

This is the part that changes how a page should be built. The retrieval step is not picking up your whole site and reading it like a brochure. It pulls sections: a paragraph, a heading with the text under it, a table of the towns you serve.

Which means a page can be good for a human and useless here. If the answer to "how much does a lift service call run" is spoken in a video, or sits three scrolls down under a testimonial slider, there is no clean passage to lift. Nothing is being penalized. There is simply nothing quotable.

Length is not the lever people assume, either. A long page with the answer buried loses to a short one that states it plainly, because retrieval is hunting for a passage that stands on its own, not a document that impresses.

The same logic argues against the everything page. A page trying to answer nine questions at once splits its own attention and rarely holds a passage worth pulling, while a page built around one real question usually does. Separate pages for each service and each town are easier to quote than one long page covering all of it.

The fix is plain writing in plain places: put the question in a heading, answer it in the first sentence underneath, and keep the specifics close by, the towns you cover, what is included, what it costs. Our AIO readiness scanner checks a page for that shape, one of more than 50 free tools we publish with no email wall.

Where RAG sits among the other AI terms

The part doing the writing is a large language model. The pages it gets handed were fetched at some point by an AI crawler. Anchoring an answer to fetched sources instead of memory is called grounding, and it is the main reason a retrieved answer beats one written from training alone.

Cutting a page into passages small enough to be retrieved on their own has a name too, content chunking, and it is worth knowing because it explains why headings do so much of the work here.