In more detail
A large language model on its own only knows what was in its training data. RAG adds a retrieval step in front of it:
- The question (sometimes rewritten or split into several searches) is run against an index, such as a web search index.
- The most relevant pages or passages are pulled back.
- Those passages are given to the model with the question, and it writes an answer based on them, often linking to the sources it used.
Google calls the same idea "grounding" when it supplies content from its Search index to Gemini. ChatGPT search, Perplexity, Claude's web search, Copilot and Google's AI features all retrieve pages in some way before they answer.
For site owners, the retrieval step looks a lot like classic search. To be used, your page has to be crawlable by the right bot, in the index being searched, relevant to the question, and easy to pull a clear passage from. Pages that bury the answer under long introductions, hide content behind scripts, or block the AI product's search crawler are less likely to be picked up.
Retrieval also explains why AI answers can change from day to day: the model may be the same, but the pages it finds change.
Why it matters when hiring an agency
When an agency talks about AI citations or AI visibility, RAG is usually the mechanism they mean, so their work should look like sound SEO: indexable pages, direct answers, good structure, accurate facts and access for AI search crawlers. Ask them to show you which retrieval-based tools they track, how often, and with which prompts, rather than accepting a single "AI visibility score" with no method behind it.