RAG

Without retrieval, a model answers from what it learned in training — a snapshot that ages from the moment it's taken. RAG bolts a search step onto generation: the engine expands your question into searches, retrieves relevant passages from a live index, and hands them to the model as context for composing the answer. The result stays current and can cite where it looked. Ask Perplexity about this month's mortgage rates and RAG is the reason the answer isn't a year stale.

For visibility work, RAG splits the problem in two. The retrieval side is winnable in weeks: if your pages are open to AI crawlers, indexed, and contain passages that survive extraction, they're candidates for every relevant retrieval from tomorrow onward. The training side — what the model already believes about your market — moves slowly and responds to your broader footprint of mentions and citations. This split is exactly the fast-path/moat-path distinction between AEO and GEO.

It also explains otherwise confusing behavior: why ChatGPT knows your firm exists but describes your old office (memory), while the same question with search enabled gets it right (retrieval); why engines differ so much in who they cite (different indexes, different retrieval logic). The discipline that ties retrieved sources to the final answer is grounding — RAG is the machinery, grounding is the contract.