Retrieval-augmented generation — when a model searches for documents first and writes its answer from what it finds, rather than answering from memory alone.
Why it is the whole game
Every AI answer that shows you sources is doing this. The system runs one or more searches, pulls back a set of documents, and composes an answer constrained by them. The model supplies the sentence structure. The retrieved documents supply the facts and the names.
The commercial consequence is blunt: a business cannot be quoted from a document that was never retrieved. There are two separate ways to lose here, and they need different fixes. Either nothing of yours came back in the retrieval step, or something came back and was too vague to quote.
Retrieval is not ranking
The retrieval runs against a query the system wrote for itself, which is often not the sentence the customer typed. One question can become several searches. So the target is coverage across the many ways a thing can be asked, not position on the one phrase you picked.
This is also why a page can be cited by an assistant while ranking nowhere for the original question, and why the reverse happens just as often.
Related: query fan-out, generative engine optimization, entity resolution.