A language model on its own answers from what it learned during training, which stops at a date and contains no page published since. With RAG, the assistant runs a search first (its own index, a partner engine, or the live web), reads the top results, and composes the answer from them. The sources it shows are those retrieved documents.
This is why GEO looks like SEO: to be in the answer, a page has to be retrieved, so it has to be indexed and relevant to the question. The writing then decides whether the model quotes it: a direct statement is easier to lift than a paragraph of context.