Retrieval Augmented Generation (RAG)

Retrieval Augmented Generation (RAG) is an inference-time architecture that pairs a large language model with an external knowledge source. When a query arrives, a retrieval system — usually semantic search over a vector index, sometimes paired with keyword search — fetches the most relevant chunks of source material. Those chunks are inserted into the model's context window as grounding, and only then does the LLM generate a response. The output therefore references real, recent documents rather than relying solely on parameters frozen at training time.

The original 2020 RAG paper from Facebook AI Research formalized the approach, but it has since become the de facto pattern for AI products that need to be accurate and current. Perplexity, ChatGPT web search, Claude's web fetch, Bing Copilot, and Google AI Overviews are all RAG systems with proprietary retrieval pipelines layered on top.

A typical RAG pipeline has four stages: (1) ingestion — content is chunked into passages, usually 100–1000 tokens each; (2) embedding — each passage is converted to a dense vector using an embedding model; (3) retrieval — a query embedding is compared to the index using cosine similarity, returning the top-k passages; (4) generation — the LLM produces an answer conditioned on the retrieved passages, often with explicit instructions to cite them.

Quality varies dramatically by implementation. Naive RAG (single-pass retrieval, top-k by similarity) handles simple factual questions well but degrades on multi-hop reasoning, ambiguous queries, and rapidly changing topics. Advanced patterns — query rewriting, hybrid keyword+semantic retrieval, re-ranking with cross-encoders, contextual compression, and agentic retrieval — close that gap but add latency and cost.

For publishers, the operational insight is that RAG turns content into a two-stage competition. Stage one is retrieval: is your page in the corpus, and does its embedding land near the user's query embedding? Stage two is citation: of the top-k passages the model sees, is yours quotable enough to actually appear in the generated answer? Optimizing for both is what distinguishes GEO from traditional SEO.