What is RAG (Retrieval-Augmented Generation)?

RAG combines two components that normally work apart: a search engine and a language model. When someone asks a question, the system does not answer directly with what the model memorized during training. It first searches an external source (a database, a set of documents, or the open web) for relevant information, then uses those results to build the answer.

The difference from a model that answers only from its training data is central. A language model without retrieval knows only what it learned up to its training cutoff and cannot state where a given fact came from, because it consults no source at the moment it answers. A model with RAG, by contrast, pulls in current information at the moment of each question, can point to specific documents, and allows its knowledge base to be corrected or expanded without retraining.

That is why RAG is associated with systems that show links alongside their answers. That ability to cite sources comes directly from the retrieval step, not from a separate mechanism.

The retrieval step works through vector comparison, not keyword matching. Both the user's question and the available documents are converted into numerical representations called embeddings, which capture the meaning of the text rather than its exact words. This makes it possible to find related content even when it uses different vocabulary than the question: a query about «reducing cart abandonment» can retrieve a document that talks about «improving checkout conversion rate» without sharing a single word. Before these embeddings are generated, documents are split into smaller pieces, a process called chunking, because comparing the question against entire documents would produce results that are too general.

The system compares the question's embedding against the embeddings of all indexed chunks and selects the closest ones, usually a few paragraphs, not entire pages. Those retrieved chunks are inserted as context directly into the prompt the model receives, alongside the original question. The model then generates its answer relying on that context, instead of depending solely on what is memorized in its parameters. This whole process runs on every new query, not just once during training, which lets the answer reflect information published minutes earlier.

Perplexity and Bing Chat show this mechanism visibly: when they answer, they cite specific links, a direct result of retrieval performed at the moment of the query, not of the model's fixed knowledge. Google applies an equivalent process in its AI Overviews. Many enterprise assistants use the same architecture over their own internal documentation, so the model can answer with company data without needing to be retrained every time a document changes.

The technical reason GEO (optimizing for AI) works is this: if a RAG system retrieves a brand's page while searching for information on a topic, that brand has a chance of being cited in the answer. If the system does not retrieve it, there is no chance at all, no matter how well that page ranks in traditional search engines.

RAG is the mechanism behind most citations in AI answers, not an optional add-on. Perplexity, Bing Chat, Google's AI Overviews, and much of the assistant landscape with access to the current web depend on a retrieval step before generating text. Content that never enters that retrieved set never reaches the generation phase, regardless of its quality.

This shifts the focus of optimization: it is no longer enough for a topic to be covered somewhere on the web, each individual page needs to be retrievable on its own. Content structured into self-contained passages, information updated with a visible date, and a source identifiable as the author directly affect whether a RAG system selects that page or discards it in favor of another. This mechanism also changes what each page competes against: two articles on the same topic no longer compete for a fixed spot in a results list, they compete to be the passage that is semantically closest to a specific user's question. These practices are not foreign to traditional SEO, but their weight changes: they stop being an incremental improvement and start deciding whether content gets to participate in the answer at all.

• Keep content crawlable by AI crawlers: do not block user agents like GPTBot, PerplexityBot, or Google-Extended in robots.txt if the goal is to be cited.
• Write self-contained passages: each paragraph should answer a complete idea without depending on the one before it, because the system retrieves individual chunks, not the whole page.
• Date and update content visibly, because a RAG system favors information it can verify as current.
• Use descriptive headings and structured data, so the retrieved chunk carries clear context on its own.
• Cover a topic with varied vocabulary, since vector search finds synonyms and different phrasings, not only the exact wording of the query.
• Cite your own sources and link to verifiable data, because RAG systems favor content that is already well-grounded.

• Blocking AI crawlers in robots.txt while expecting to be cited in answers: these are incompatible goals.
• Writing very long paragraphs that mix several ideas, which makes it harder for the system to retrieve the exact chunk that answers a question.
• Leaving content without a visible publish or update date, which leads the system to assume it may be outdated.
• Confusing RAG with a ranking trick separate from SEO or GEO, when in fact the same optimization practices serve both.
• Optimizing only the homepage or main page, forgetting that each page is retrieved independently.