Retrieval Augmented Generation (RAG)

What RAG means

A large language model's core knowledge comes from the data it was trained on, fixed at a point in time and baked into its parameters. RAG adds a second step in front of the generation itself: before writing an answer, the system searches an external source, an index of web pages, a document store, or a company's internal knowledge base, retrieves the passages most relevant to the question, and feeds them to the model as additional context alongside the question. The model then generates its answer using both its trained knowledge and the retrieved material, typically favoring the retrieved content for anything specific or current. The term comes from a 2020 research paper by Meta AI researchers, who proposed combining a retrieval step with a generative model to reduce factual errors and let a model answer questions about information it was never trained on. Search-connected AI assistants such as Perplexity, and web-search-enabled modes of ChatGPT, Claude and Gemini, are all practical examples of RAG in production: the live search step is the retrieval half, and the written response is the generation half.

How it works

A RAG system has two main components working in sequence. The retrieval component takes the user's query, searches an index, which might be a live web search, a vector database of embedded documents, or both, and returns the passages judged most relevant, typically ranked by a similarity score. The generation component, the language model itself, then receives the original question along with these retrieved passages as added context and produces an answer that draws on both. This differs from simply searching and showing raw results, since the model synthesizes the retrieved material into new, direct prose rather than presenting a list of matching documents. It also differs from a model answering purely from its training data, since the retrieved passages can include information published well after the model's training cutoff, which is one of the main problems RAG was designed to solve. The quality of a RAG system's output depends heavily on the retrieval step: if the search returns irrelevant or low-quality passages, the model's answer inherits that weakness even if the model itself is capable, since it is working from what it was actually given rather than an idealized version of the web. Citations shown alongside an AI-generated answer typically map directly to the passages the retrieval step pulled in, which is why RAG-based answers can attribute sources in a way a model answering from training data alone cannot.

Why it matters for AI visibility

RAG is the mechanism behind most of the citation-based visibility a brand can actually earn in AI-generated answers. When an assistant like Perplexity or a web-search-enabled ChatGPT session cites a source, that citation traces back to the retrieval step of a RAG system pulling in a specific page because it matched the query well. This means a brand's content has a real, mechanical path to being surfaced, being written clearly enough and structured well enough to get selected during retrieval, rather than depending entirely on whatever happened to be in a model's training data at some earlier point. Understanding RAG explains why technical accessibility, being crawlable and indexable by the systems doing the retrieving, matters just as much for AI visibility as the writing itself: a page invisible to the retrieval step can never be pulled into a generated answer, regardless of its quality. This is a practical, actionable insight rather than an abstract one: a marketing team cannot change how a model was trained, but it can absolutely change whether its own pages are crawlable, well-structured and specific enough to win the retrieval step that decides what gets cited.

Good practices

• Keep content crawlable and indexable by the AI crawlers and search systems that feed retrieval steps, since unreachable content can never be retrieved.
• Write passages that stand on their own and answer a specific question clearly, since retrieval selects and extracts fragments, not entire pages.
• Include specific, current details, since RAG's main advantage over training-data-only answers is surfacing information a model was never trained on.
• Structure content with clear headings that match likely queries, which helps a retrieval step match the right passage to the right question.
• Test whether your content actually gets retrieved by asking the real questions a RAG-based assistant's users would ask.

Common mistakes

• Assuming every AI assistant uses RAG, when some modes answer purely from training data with no live retrieval step at all.
• Writing content that reads well as a whole page but has no single passage that stands on its own for a retrieval system to extract.
• Ignoring technical crawlability, which blocks the retrieval step regardless of how well the content itself is written.
• Confusing RAG with a search engine, when RAG specifically adds a generation step that synthesizes retrieved passages into new text.

Related terms

• Grounding: the broader practice of anchoring an answer to retrieved sources, which RAG is one method of achieving.
• Knowledge Cutoff: the training data limit RAG is specifically designed to work around through live retrieval.
• Answer Engine: the category of system, such as Perplexity, that commonly runs on a RAG architecture.
• AI Overview: a Google feature that functions as a RAG system, retrieving indexed pages before generating its summary.

Frequently asked questions