RAG (retrieval-augmented generation)

RAG is a way of building AI answers where the system first finds relevant pages or documents, then has the language model write its answer from them.

AI searchUpdated 2 min read

Also called retrieval-augmented generation

What it means

RAG stands for retrieval-augmented generation. It has two steps:

  1. Retrieve: search for pages or documents that match the question.
  2. Generate: give those pages to a language model and have it write an answer from them.

The idea comes from a 2020 research paper by Facebook AI researchers. Today it's how almost every AI search tool works. Google calls the same idea "grounding": tying the answer to real sources.

Why AI search uses it

A language model on its own only knows what was in its training data, which stops at a certain date. RAG lets it answer with current information, and it gives the tool sources to show.

What it means for brands

With RAG, the answer can only include what's in the pages the system retrieved. That's the core of AI search for businesses: if you're not on the pages the AI reads, you're much less likely to be in the answer.

Our study measured this directly. 90% of the brands AI recommended were on a "best X" list it had retrieved for that answer. Brands on none of the lists were recommended 2.5% of the time.

The retrieved pages also change between runs. Only 42% of cited pages were the same from one day to the next, which is a big reason answers change so much.

Example

Ask Perplexity "which CRM has the cheapest paid plan right now?". It searches the web, pulls the pricing pages and a recent comparison, and writes an answer with today's prices and links to those pages.

Without retrieval, the model could only guess from what it learned in training, which may be a year or more old.

Frequently asked questions

Does every AI answer use RAG?

No. Some questions get answered from training alone, without a search. Buying questions usually trigger a search, because products and prices change.

What does RAG change for my brand?

It means the pages AI reads today matter more than what it learned in training. You can change those pages, or get onto them, now.

Why does RAG make answers change so much?

Each run can search differently and get different pages back. In our study, only 42% of cited pages were the same from one day to the next.

Can I tell when an answer used RAG?

Usually, yes: it shows sources or links. An answer with no sources was probably written from training data alone.

Go deeper

All posts →
Research

AI product recommendations are mostly a summary of listicles

We fetched every 'best X' page ChatGPT, Gemini and Google AI cited in 839 answers. 90% of the brands they recommended were on one of those lists. Brands on none were recommended 2.5% of the time. Here's how it works and what to do about it.

Make AI search your next revenue channel

Start free trial