What it means
The term comes from a 2020 research paper that combined a pre-trained language model, its parametric memory, with a searchable index of documents, a non-parametric memory. Most answer engines work this way when they search: they retrieve pages, pass selected passages to the model and generate an answer that may cite them.
Google uses 'grounding' for the same pattern in Gemini Apps and in Grounding with Google Search on Vertex AI. Retrieval makes answers more current and easier to check, but the model still decides what to say and which sources to show.
Why it matters
RAG is why AI visibility is not frozen at training time. When an engine retrieves at answer time, the pages it can find today, including your documentation and third-party reviews, can change tomorrow's answer. It is also why crawler access matters.
Example
A buyer asks an assistant whether a project management tool includes Gantt charts on its free plan. The assistant retrieves the vendor's pricing page and a review from last year, and the answer repeats the outdated review. Updating the pricing page helps only if the engine retrieves it and weighs it above the review.
Common misunderstandings
- RAG does not guarantee accuracy; the model can misread or over-generalise a retrieved page.
- Not every answer uses retrieval. Without it, the answer rests on what the model learned in training.
- Being retrieved is not the same as being cited or recommended.