What Is Retrieval-Augmented Generation (RAG)?

How RAG Works

A retrieval-augmented system answers a question in three moves: retrieve, augment, generate.

First it retrieves. The user’s question is turned into a query, and a retriever searches a collection of documents for the passages most relevant to it. Modern systems do this with vector search: both the query and the stored passages are converted into embeddings — numerical vectors that place similar meanings close together — and the retriever returns the passages whose vectors sit nearest the query’s. Because the match is on meaning rather than exact words, a query about “raising a newborn’s sleep schedule” can pull a passage that never uses those exact terms.

Then it augments the prompt. The top retrieved passages are pasted into the prompt alongside the original question, usually with an instruction like “answer using the following sources.” The model now has the relevant facts sitting in its context window instead of having to recall them from training.

Finally it generates. The language model writes an answer conditioned on the augmented prompt. Because the answer is grounded in specific retrieved text, the system can attribute claims back to their source — this is why AI answers can show citations at all. The passages that got retrieved become the grounding sources the answer is built from.

The important consequence: the model never invents which documents exist. It can only work with what the retriever handed it. If your content is not in the retrieved set, the generator cannot cite you no matter how good the page is.

Components of a RAG System

Every RAG pipeline has three core parts.

The retriever finds candidate passages. Early systems used keyword search; today most use dense retrieval, where an embedding model scores passages by semantic similarity to the query. The retriever is the gatekeeper — it decides which handful of passages, out of millions, the generator is even allowed to see.

The vector store holds the corpus. Before anything can be retrieved, documents are split into chunks (see chunking), each chunk is embedded, and the embeddings are stored in a vector index built for fast nearest-neighbor lookup. How a document is chunked directly affects what can be retrieved: a self-contained paragraph is retrievable as a clean unit, while an answer smeared across three sections may never surface as one coherent passage.

The generator is the language model that writes the final answer from the augmented prompt. It supplies fluency and reasoning; the retriever supplies the facts. Swapping the corpus changes what the system knows without retraining the generator — which is exactly why RAG became the default way to give general-purpose models access to fresh or private knowledge.

Example of RAG

The clearest worked example is the paper that named the technique. In 2020, Patrick Lewis and co-authors at Facebook AI Research and University College London published “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (submitted May 22, 2020). They combined two things that had usually been kept separate: a language model’s “parametric memory” — knowledge baked into its trained weights — and “non-parametric memory,” an external index it could look things up in.

The build was concrete. The retriever was a dense neural retriever (Dense Passage Retrieval) that searched a vector index of Wikipedia, roughly 21 million passages. The generator was BART, a pre-trained sequence-to-sequence model. For each question, the retriever pulled relevant Wikipedia passages, those passages were fed to BART, and BART generated the answer. The whole thing was trained end-to-end, treating the retrieved document as a hidden variable the system learned to pick well.

They proposed two flavors that still describe RAG systems today: RAG-Sequence, which conditions the entire answer on one retrieved passage, and RAG-Token, which can draw on different passages for different words of the answer. On open-domain question answering — where a system must answer factual questions without being handed the relevant document — the RAG models set a new state of the art on three benchmark tasks, and produced answers that were more specific and better grounded than a comparable model relying on trained-in memory alone.

The mechanism is the whole point: the facts lived in the external Wikipedia index, not in the model’s weights. Update the index and the system’s knowledge updates, no retraining required. That decoupling — a fixed generator over a swappable, searchable corpus — is the pattern every production RAG system inherits.

Why RAG Matters for SEO and GEO

Here is the reframe that makes RAG the most important term in this glossary for anyone doing SEO: AI search is a RAG system running over the web, and your pages are its corpus.

When ChatGPT browses, when Perplexity answers, when Google assembles an AI Overview, the same three moves run. A retriever searches an index — for Google, the same index that powers blue links, often expanded through a query fan-out that issues several background searches per question. The top passages are pulled into the model’s context. The model generates a cited answer from them. Swap “Wikipedia” for “the open web” and the 2020 architecture is exactly what sits under Google’s AI Mode and every AI answer engine.

That changes where your leverage is. You cannot touch the model’s weights, and you cannot argue your way into an answer. What you can control is the retrieval step — whether your content gets fetched — and the quality of the passage that gets lifted. Three things follow directly:

• You must be retrievable. If AI crawlers cannot fetch your page, or it never enters the index, no amount of good writing matters. You are not in the corpus.
• You must be chunkable. Retrievers work on passages, not whole pages. A self-contained paragraph that answers one question cleanly is a retrievable unit; a good answer scattered across a long page is not. This is why chunking-friendly structure — clear headings, one idea per block — wins retrieval.
• You must be quotable. Even after retrieval, the generator reaches for the passage it can lift with least risk: a plainly stated, self-standing claim. This is the entire premise of generative engine optimization — writing passages an engine will actually use.

Understand RAG and the rest of AI-search strategy stops being a grab bag of tactics. It becomes one clear objective: get into the retrieval corpus, and be the cleanest passage the generator can quote.

Frequently Asked Questions

The Bottom Line

RAG is the architecture behind grounded AI answers: retrieve relevant passages, then generate a response conditioned on them. For anyone doing SEO or GEO, the key realization is that AI search IS a RAG system over the open web, and your content is the retrieval corpus. You cannot edit the model, but you decide whether your passages are clean enough to be fetched and quotable enough to be used. Optimize for the retrieval step.