Retrieval-Augmented Generation (RAG) is the trick behind most AI answers: fetch relevant, current documents the moment a question is asked and let the model answer from those, rather than from memory alone.
Left to itself, a language model answers from training data, a snapshot frozen months ago that knows nothing about your latest price or your newest page. RAG closes that gap. A question comes in, the system first goes and retrieves documents that look relevant (usually by semantic search), and only then hands those passages to the model to write from. The citations stacked under a Perplexity or AI Overview answer are exactly the documents that retrieval step pulled.
This is why being crawlable and cleanly structured suddenly matters so much. A page that can't be retrieved can't be cited, and the model will just answer from training data or from a competitor's page that could be reached. RAG is what turns "is my content accessible and parseable" from a technical footnote into the price of admission for AI answers at all. It's the mechanism that makes GEO and AEO worth doing.
What makes a page retrievable
A page earns its way into that retrieval step by being reachable in the first place, crawlable and not hidden behind script the crawler can't run. It has to read clearly enough that meaning-based search matches it to the right questions, carry enough structure that the right passage can be pulled out on its own, and stay current, since a live page beats a stale training memory every time. Ask an engine a pricing question and it will retrieve your live pricing page and quote today's number, as long as that page was reachable when it went looking.
Related terms
- Semantic Search: the method RAG usually uses to find passages
- Structured Data / Schema Markup: keeps retrieved passages parseable
- Generative Engine Optimization (GEO): the work of being retrievable in the first place