Extended definition
Technique where an LLM accesses external data (web search, databases) before generating the answer. Perplexity and ChatGPT with browsing use RAG.
Context and application
RAG is how LLMs “solve” the knowledge cutoff problem: instead of relying only on training data, they do live search (Bing API, Google Search API, proprietary indexes) and use the results as context for generation. For SEO/GEO: if the LLM uses RAG with Google Search as source, then your classic SEO ranking matters directly to be cited in the AI answer. If it uses another engine (Perplexity has its own indexer), tactics are partially different. Google AI Overviews uses a combination of training data + RAG on Google index — meaning GEO + classic SEO = partially the same optimizations.