Having a language model write its answer from documents retrieved at question time rather than from memory alone; the method behind the citations you see in AI answers.
What is RAG?
RAG stands for retrieval-augmented generation. A plain language model answers from its training data, which is frozen at some date and whose contents the model cannot precisely account for. RAG inserts a step: when a question arrives, relevant documents are searched for and fetched first, and the model then writes its answer against them.
The effect runs two ways. Answers become current — information published after training can still appear. And answers become traceable: the source links under a generated answer usually come out of that retrieval step. The citations you see in AI interfaces are not what the model remembered, but what it just read.
What it means for brand visibility
RAG explains why GEO is content work rather than a technical trick. The model does not need to have "learned" anything about you; it only needs to find you at the moment the question is asked. The goal is not getting into training data — it is being clear and reachable enough to be picked in the retrieval step.
- Retrieval usually runs through search and crawling; a site closed to AI crawlers is not in that race at all.
- What gets quoted is not the whole page but the passage answering the question — which is exactly why clear definitions and FAQ blocks do well.
- If your information is stale, the model may relay it as current; outdated prices and discontinued product pages are a risk in themselves.
- The same method underpins how tool-using AI agents look at your site.
Frequently asked questions
Does RAG stop models from getting things wrong?
It reduces the risk rather than removing it. Grounding an answer in real documents makes invention less likely, but if the retrieved source is wrong, stale or off-topic, the model will relay it confidently. Quality is capped by the quality of what was retrieved — which makes the freshness of your own pages directly consequential.
What changes for my brand in practice?
The target is not getting into training data but being retrievable and quotable at question time. That means a crawlable site, brand facts written clearly and consistently, answers given directly on the page, and stale information cleaned up. All of it is ordinary good content and good technical hygiene.
Related concepts
AI Agents
A system where a language model plans and carries out a multi-step goal by using tools; it does not just produce text — it searches, reads pages and takes actions.
GEO (Generative Engine Optimization)
The discipline of making sure a brand is mentioned accurately, currently and favorably inside answers generated by ChatGPT, Perplexity, Google's AI summaries and other generative engines.
LLM (Large Language Model)
An AI model trained on very large volumes of text that generates language by predicting what comes next; the technology behind products like ChatGPT, Claude and Gemini.