What is Retrieval-Augmented Generation (RAG)?
The architecture where an LLM retrieves relevant documents first, then generates its answer from them — the engine of citable AI search.
Definition
Retrieval-augmented generation is the architecture behind grounded AI answers: a retrieval step finds relevant documents (from a web index or database), and the model generates its answer conditioned on those documents, citing them. ChatGPT Search, Perplexity, AI Overviews, and enterprise AI search all run RAG variants.
Why it matters
RAG is why GEO works at all. Training-data influence is slow and diffuse, but the retrieval step is a search problem happening every time someone asks — with winners chosen in milliseconds from the candidates the index returns. Understanding RAG tells you exactly where to compete: be in the index (crawlable, indexed), be retrieved (relevant, authoritative for the prompt), be used (quotable, data-rich passages the generator selects).
Frequently asked
How does RAG choose documents?
A pipeline of query interpretation (often with fan-out), index lookup, semantic ranking via embeddings, and passage selection. Each stage prunes candidates — you must survive all of them to be cited.
Does RAG use my whole page?
No — it extracts passages. That's why self-contained, answer-shaped sections matter more than overall page length: the unit of retrieval is the chunk, not the URL.
Related terms
- Rich ResultsEnhanced search listings powered by structured data — stars, FAQs, products — and a proving ground for machine-readable content.
- robots.txtThe file governing which bots may crawl what — now the control panel deciding your participation in AI training and AI answers.
- AI CitationsThe linked sources an AI assistant credits when generating an answer. Being cited is the AI-era equivalent of ranking #1.
- AI CrawlerBots operated by AI companies — GPTBot, ClaudeBot, PerplexityBot, Google-Extended — that fetch web content for model training or live answer retrieval.