What is Retrieval-Augmented Generation (RAG)?
Retrieval-augmented generation (RAG) is a technique where a language model retrieves documents from an external source at the time a question is asked, then writes its answer from those documents rather than from its training data alone. The retrieved documents are what the model can cite.
A model without retrieval answers from parameters fixed at training time. It cannot know about anything published since, and it has no document to point at, which is why an ungrounded answer carries no links.
With retrieval, the question is first turned into a search — against a web index, a vector store, or both. A small number of documents come back, usually between three and ten, and those documents are placed into the model's context alongside the question. The answer is written from that context.
This is the mechanism that makes AI visibility a measurable thing rather than a metaphor. Being cited is not a matter of the model "knowing" a brand; it is a matter of a page being retrieved into that context window at answer time, then being quotable enough to survive the summarisation that follows.
It also explains why the retrieval step and the writing step fail differently. A page can be retrieved and then not cited because nothing in it was extractable, or never retrieved at all because it did not match the query. Those are different problems with different fixes, and a visibility tool that reports only "not cited" cannot tell them apart.
Why Retrieval-Augmented Generation matters
Citation is decided at retrieval time, not by reputation. A page that is never retrieved cannot be cited no matter how authoritative the brand behind it is.
Only retrieved documents can be cited
The candidate set is typically 3-10 documents. Everything outside it is invisible to that answer regardless of quality.
Retrieval matches passages, not pages
Systems commonly chunk a page and match a section. A single strong self-contained passage can win an answer for an otherwise unremarkable page.
Freshness is a retrieval signal
For queries with recent intent, retrieval favours recently published or updated documents, which is one of the few levers that works quickly.
Two failure modes, two fixes
Not retrieved is a findability problem: matching the query wording and topical coverage. Retrieved but not cited is an extractability problem: format and self-containment.
Common questions
No. Search returns a list of results for a person to read. RAG uses a search step internally, then writes a single answer from what it retrieved. The search is a means to an answer, not the output.
Only for what it retrieves. The underlying model is still fixed at its training cutoff; retrieval supplies current documents on top of it. If retrieval fails or returns nothing relevant, the answer falls back to training data with no indication to the user.