Grounding & RAG (retrieval-augmented generation)
Also known as: retrieval augmented generation, grounded answers
The technique that makes an assistant search the live web, read what it finds, and write its answer from those pages — which is why citations exist at all.
Last revised 2026-09-07. Also available as markdown — request this URL with Accept: text/markdown.
In plain English
A language model on its own writes from what it absorbed during training: broadly knowledgeable, frozen in time, and unable to tell you where anything came from. Grounding bolts a search step onto the front. The assistant looks things up, reads the results, and answers from them.
This is the single most useful thing to understand about AI visibility, because it means your current pages matter. If answers came only from training data, the only way in would be to have been famous before the cut-off. Because they are grounded, a page you publish this week can be quoted next week.
How to fix / set it up
- Make the fact you want quoted exist as plain text in the HTML, not only in an image or a chart.
- Give each page one clear subject and a heading structure that matches the questions it answers.
- Keep pages reachable without JavaScript execution where you can — several retrievers read raw HTML only.
- Publish updates as dated revisions rather than silent edits, so freshness is visible.
Full guide: why AI cannot read JS-only pages →
The technical detail
The pipeline is: rewrite the question into search queries, retrieve candidate documents, select passages, and generate an answer conditioned on those passages with citations attached to the claims they support.
Retrieval is usually hybrid — keyword matching plus vector similarity over embeddings — so both the words on the page and its meaning matter. See embeddings-semantic-search.
The retrieval step is where you are won or lost, and it is unforgiving of packaging. A page whose relevant fact sits in an image, a script-rendered widget or a PDF behind a login can be perfectly authoritative and still never be retrieved.
FAQ
If answers are grounded, does model training data still matter?
It matters for whether the model recognises your brand as an entity at all, which shapes how confidently it names you. Grounding decides what is quoted today; training shapes the priors.
Why was I cited for a page that does not rank on Google?
Retrieval and ranking are different selections. A page can be an excellent passage-level match for one question while ranking poorly on the broader query.