How AI answers are made
Retrieval-augmented generation (RAG)
Also called: RAG
Retrieval-augmented generation (RAG) is a way of building AI answers in two steps: first find relevant text (retrieval), then let a language model write the answer using that text (generation).
In plain words
Think of an open-book exam. The model doesn’t have to remember everything, because it can look things up first.
The name comes from a 2020 research paper by Lewis and others. They combined a language model with a searchable index of Wikipedia.
Example
A company builds a help chatbot on top of its own manuals. A customer asks how to reset a device. The system first finds the most relevant manual paragraphs, and the model then writes a short answer from them.
Why it matters for AI search
Google says its AI features use RAG. They retrieve relevant, up-to-date pages from the Search index and generate a response with clickable links to those pages.
In a RAG system, only text that gets retrieved can end up in the answer. If your page is not found in that first step, it cannot be cited.
What you can do
- Write sections that answer one question clearly, so a retrieved passage makes sense on its own.
- Use the words people actually use for a topic, next to any technical terms.
- Keep your pages crawlable and indexable: for Google’s AI features, retrieval starts from the Search index.
Often confused with
- Grounding
- Google’s guide describes RAG as "also known as grounding". In practice the two words point at the same idea.
- Fine-tuning
- Fine-tuning trains the model further, which changes the model itself. RAG leaves the model as it is and hands it text to read.
Sources
Related terms
How does your page do in AI search?
GEO-Score checks one page on 22 signals, from AI crawler access to structured data, and shows what to fix first.
Analyze a page