Definition
RAG (retrieval-augmented generation) is the pattern in which an AI system searches for relevant documents at the moment a question is asked and writes its answer from those documents, usually with links to them. In AI search, RAG is what turns a chatbot into an answer engine: ChatGPT search, Perplexity, Google AI Overviews and AI Mode, and Microsoft Copilot all retrieve web content before they respond to many questions.
This entry covers RAG as it appears inside AI search products. The technique itself, its 2020 research origin and its components are described under retrieval-augmented generation.
The vendors describe the same basic loop. OpenAI's help center says ChatGPT 'may search the web automatically when your question would benefit from current information', rewrites the prompt into 'one or more targeted queries' for its search partners, and may send more specific follow-up queries after reading the first results (OpenAI). Google says AI Mode breaks a question into subtopics and issues many queries at once, a technique it calls query fan-out. In practice the loop has four steps:
- Decide whether to retrieve at all. Many prompts are answered from parametric knowledge with no search, which is what grounding rate measures.
- Turn the prompt into search queries, called grounding queries.
- Fetch candidate pages from an index, split them into passages, and rerank those passages so only the best few reach the model.
- Write the answer from the selected passages and attach citations.
The scale is large. Google said at I/O in May 2026 that AI Overviews had more than 2.5 billion monthly users and AI Mode more than 1 billion (Google). OpenAI said on September 29, 2026 that ChatGPT had 1.2 billion weekly users. Each answer that triggers retrieval is a small selection contest in which a handful of sources win.
For GEO, RAG sets the rules of that contest. A page can only be cited if the system can fetch it, if it appears in the results for the queries the model wrote, and if one of its passages survives ranking. That has practical consequences. Crawler access comes first: OpenAI states that sites must allow OAI-SearchBot to be eligible for ChatGPT search results. Ranking in the underlying search results for the rewritten queries comes second, which is why how ChatGPT searches the web matters more than how a user phrases a prompt. Passage quality comes third: a clear, self-contained answer with specific facts beats a long page that only answers the question in aggregate.
Teams measure RAG outcomes by tracking a fixed set of prompts across engines, recording which URLs are cited, and comparing them with the queries each engine ran. When a competitor is cited and you are not, the usual cause is visible in that chain: the model searched for something you do not rank for, could not fetch your page, or found a sharper passage elsewhere.
Examples of RAG (Retrieval-Augmented Generation)
- A user asks ChatGPT which payroll tools integrate with Xero for a 15-person agency. ChatGPT searches for 'payroll software Xero integration small business' and cites the four vendor and review pages that ranked for that query.
- A SaaS company accidentally blocks OAI-SearchBot in robots.txt. Its pages stop appearing as ChatGPT search citations, although ChatGPT still names the brand from training data when no search runs.
- Perplexity answers a question about a product recall announced that morning by citing two news reports, information no model could have learned during training.
- A GEO team runs the same buying prompt in Google AI Mode and ChatGPT and gets different cited sources, because each product writes its own queries and retrieves from its own index.
