Retrieval-Augmented Generation (RAG)

RAG is the reason AI visibility is winnable at all. If assistants only answered from training data, changing what they say about a business would mean waiting for the next model. Because they retrieve live, the answer changes when the evidence changes — update the pages, profiles, and reviews the retrieval step finds, and the generated answer follows.

The pipeline has three stages, and a business can be eliminated at any of them. Retrieval finds candidate documents, usually via a search query the assistant composes itself. Selection narrows those to the few that fit the context window. Generation writes an answer from what survived. Ranking well for the query the assistant happened to compose gets you into stage one; being clearly and quotably relevant gets you through stage two; nothing gets you through stage three except being in the text the model is reading.

The compression at stage two is more brutal than most people picture. A search might return ten blue links, of which perhaps three or four pages are actually read in enough depth to influence the answer. That is the real shortlist, and it is decided before any evaluation of which business is better.

This also explains why the answer changes when nothing about your business did. Retrieval is sensitive to phrasing, location, timing, and whatever the assistant decided to search for. A competitor publishing a strong comparison page, or a forum thread gaining traction, can displace you from the retrieved set entirely — you were not outranked, you were simply not fetched.

It is why what gets retrieved is the competitive question. When an assistant answers "best roofer in Phoenix," a handful of retrieved pages decide the shortlist. Our research records exactly which sources those are, trade by trade, which is the closest thing to a map of this stage that exists.