Retrieval-Augmented Generation is the process an AI system uses to fetch relevant sources before writing an answer, rather than answering from memory alone. When ChatGPT searches the web or Perplexity pulls pages, that is RAG. It matters because being retrieved is the step before being cited.
The short version
- A model without retrieval answers from training data: fixed, undated, uncitable.
- RAG adds a fetch step, so the answer is grounded in real sources that can be named.
- Two gates: your content must be retrieved, then it must be selected for the answer.
- Citations only exist because of retrieval. No retrieval, no link, no attribution.
- Optimising for retrieval and optimising for selection are different jobs.
How it works, without the jargon
A language model on its own is a fixed snapshot. It has read an enormous amount up to a point in time and can produce fluent text from that, but it cannot tell you what changed last week and cannot honestly attribute a claim to a source, because it is not reading anything as it answers.
Retrieval-Augmented Generation adds a step in front. The system converts your question into one or more searches, fetches a set of candidate documents, and puts that material in front of the model along with your question. The model then composes an answer grounded in those documents, and can cite them because it genuinely used them.
This is why answers now carry links at all, and why the same question can produce different answers on different days.
For anyone trying to be cited, the important consequence is that there are two separate gates.
Gate one is retrieval. Your page has to be findable and fetchable by the retrieval step. If crawlers are blocked, the content needs JavaScript to render, or the page does not surface for the underlying query, you never enter the candidate set. Nothing else you do matters.
Gate two is selection. Being retrieved is not being used. From the candidate set, the model picks what to actually cite, and it favours material that answers directly, states things specifically, and comes from a source it can identify and trust.
Most businesses that fail are failing at gate one and assume they are failing at gate two, so they write more content when they should be fixing access and structure.
How to check it yourself
Ask Perplexity a question your content should answer, then look at the sources it lists. Those are the pages that passed both gates. If your competitors are there and you are not, look first at whether your page is reachable and readable at all, then at whether it contains a liftable answer.
What to do about it
- Confirm the crawlers can reach you. Check
robots.txtand confirm no firewall or CDN rule is quietly blocking AI user agents. Our robots.txt guide for AI crawlers covers the specific names. - Render main content server-side. If the text needs JavaScript, assume some retrievers will not see it.
- Write the answer first, so the selection step has something clean to take.
- Be specific and dated, because grounded answers favour material that commits to facts.
- Publish a machine-readable summary of who you are, so the model can attribute confidently.