Retrieval-augmented generation (RAG)
Retrieval-augmented generation (RAG) is a technique where an AI system fetches relevant documents, often from a live web index, and passes them to a language model to ground its answer. Instead of relying only on what the model memorised during training, the answer is built from content retrieved at the moment you ask. For marketers, it means an AI assistant's description of your brand can come straight from web pages, including yours.
The process has two stages. When you ask a question, the system first searches an index for relevant material, then feeds the best passages to the model alongside your question, and the model writes its answer with those passages in front of it. That is why grounded AI answers often include citations: the sources are not decoration, they are the raw material the answer was assembled from.
This matters for marketers because retrieval is a filter. The model can only quote and cite what the retrieval step surfaces, so if your pages are never retrieved for a query, you are invisible in that answer no matter how good your content is. Retrieval tends to favour pages that address the question directly in self-contained passages, which is why so much GEO advice comes down to answering a specific question cleanly near the top of the page.
Not every AI answer uses RAG. When an assistant replies without searching, it works from training data alone, which is a snapshot that may be months or years out of date. Grounded modes reduce that staleness but do not remove it, and providers do not publish the full details of their retrieval pipelines. Each engine also decides differently when to search at all, so the same question can be grounded in one session and answered from memory in the next.
The practical takeaway is that you have two jobs. Make your content retrievable, which means it must be crawlable and present in the indexes the engines draw on. And make it quotable, so that when a passage is pulled into the model's context it survives as the sentence the answer is built around.