Retrieval-augmented generation (RAG): definition

Retrieval-augmented generation is a model answering with documents it was handed at answer time, not only with what it learned in training.

Without retrieval, a model answers from parametric knowledge. That knowledge is frozen at a training cutoff. With retrieval, the system runs a search, pulls back a set of documents, and passes their text into the model's context window. The model then writes an answer grounded in that text and may cite some of the sources.

The word "augmented" is the point. Retrieval adds fresh, specific material to a model that would otherwise guess from memory.

A typical AI search answer runs through these steps:

1. Query rewriting. The system turns the user's prompt into its own search queries. OpenAI says ChatGPT search typically rewrites the query into one or more targeted queries, and Google says it may use a query fan-out across subtopics.
2. Retrieval. Those queries run against an index. The system keeps a shortlist of candidate documents.
3. Context assembly. The shortlisted text goes into the model's context, sometimes split into pieces by the operator.
4. Generation. The model writes the answer and attaches citations.

Two caveats matter. First, retrieval only reaches pages a crawler could fetch and index earlier, so access comes before everything. Second, as of October 2026, retrieval is discretionary: ChatGPT decides per query whether to search, and Google AI Overviews often do not trigger. Some answers skip retrieval entirely.

Separating the stages shows where GEO evidence actually sits. Nearly every published GEO effect size was measured on documents already inside the model's context, with no live index call. The original GEO paper supplied sources to an engine the authors built. C-SEO Bench did the same and found most conversational SEO methods "largely ineffective" and often harmful to ranking.

A 2026 critical survey of 45 GEO studies (arXiv 2607.14035) summed it up: already-retrieved content can change its citation, but no reviewed technique shows a stable, cross-platform effect on discoverability. In plain terms, those studies answer "once you are in the context, what happens?" They do not answer "how do you get in?" You can read the fuller picture in our guide to generative engine optimization.

The next move is to check what you control before retrieval: whether AI crawlers can fetch your pages, whether your key facts appear in HTML before JavaScript runs, and whether your pages are indexed with snippets allowed. Nobody outside the operator controls which candidates get retrieved or in what order.

• "Chunk your content for RAG." Chunking is done by the retrieval operator inside its own index. The document chunking study measured that operator-side choice, which publishers cannot set.
• "Cited means used." OpenAI notes the sources consulted often outnumber the citations. The mechanistic study of how LLMs cite repaired over 90% of missed citations by adjusting model internals without changing the sources.
• "Every AI answer is retrieved." Retrieval fires per query. A single score that mixes retrieval-on and retrieval-off answers averages two different systems, which is why AI visibility scores need careful reading.
• "GEO adds up to 40% visibility." That figure came from a lab engine with sources already in context.

• Semantic search: retrieval by meaning rather than exact keywords, often the retrieval step inside RAG.
• Faithfulness: a RAG metric for whether an answer's claims are supported by the retrieved text.
• Grounding: the general idea of tying a model's answer to external sources, of which RAG is the common implementation.