Retrieval-Augmented Generation (RAG) combines a search step with a language model. When a question comes in, the system retrieves relevant passages from an index, such as the web or a company's documents, and passes them to the model along with the question. The model writes an answer based on those passages. Most AI search products, including ChatGPT search, Perplexity and Google's AI features, use some form of this pattern.
How a RAG pipeline works
The details vary by product and are rarely published in full, but the general shape is well understood.
- Query processing: the question is rewritten or split into several search queries.
- Retrieval: a search index returns candidate documents or passages, often using keyword matching, vector similarity, or both.
- Ranking and selection: candidates are reordered and a small number are kept to fit the model's context window.
- Generation: the model writes an answer using the selected passages as evidence.
- Attribution: the system links parts of the answer back to the sources used, shown as citations.
Why RAG matters for search visibility
A model's training data is fixed at a cutoff date and may never have included your pages. RAG is how current information, such as a new pricing plan or a recent integration, gets into an answer. For a brand, that means the retrieval step is the door: if your page is not retrieved for the query, the model does not see it.
Retrieval usually works on passages rather than whole pages. A page that contains one clear, self-contained passage answering a specific question is easier to retrieve and use than a page that spreads the answer across several sections.
What this suggests for content
Because the inner workings of commercial systems are not public, treat these as reasonable inferences, not rules.
- Be retrievable: crawlable, indexed and relevant to the words people use.
- Be quotable: put the answer in a passage that makes sense on its own.
- Be specific: concrete facts, names, numbers you can support and clear scope give the model something to cite.
- Be current: retrieval systems can prefer fresher content for time-sensitive queries.
RAG inside SaaS products
Many SaaS companies also build RAG into their own products, for example a support assistant that answers from the help center. The same lessons apply in reverse: a help center written in clear, self-contained articles makes your own assistant more accurate, and makes your docs more useful to outside AI tools too.
Example
A fictional data tool's changelog
Querybase, a made-up data warehouse tool, launched native dbt support two months ago. A model answering from training data alone would not know this. When a user asks Perplexity “does Querybase support dbt?”, the system searches the web, retrieves Querybase's integration page and changelog, and answers using them. If those pages did not state the fact plainly, the answer could be wrong or vague.
Common mistakes
Assuming the model “knows” your product
Training data is old and patchy. For current facts, AI answers depend on retrieving your pages or pages that mention you.
Writing for retrieval at the expense of readers
Chopping pages into disconnected fragments makes them harder for people to read. Clear structure serves both.
Questions
They are closely related. Grounding is the goal of tying an answer to verifiable sources; RAG is one common way to do it.
Many retrieval systems combine vector similarity with keyword matching, but specific products rarely document their exact methods.
