What is Grounding?
- Grounding
- Grounding is the practice of tying a model's generated answer to specific retrieved documents, so each claim traces back to a source rather than to the model's own parameters. Grounding is what makes a citation possible, because an assistant can only attribute a sentence to a page when the sentence was built from that page. An ungrounded answer may still be correct, but nothing in it can be credited to anyone.
01 In practice
Grounding is why parseability decides more than prose quality does. A model grounding an answer works from the text of the response it fetched, so a page that assembles its content in the browser hands the retrieval agent an empty shell to ground against.
Self-contained passages give grounding something to attach to. A sentence that names its own subject and states one checkable claim can be lifted and attributed; a sentence whose meaning depends on the heading above it and the paragraph before it does not survive the trip.
Ungrounded answers are where a brand is most often described wrongly. With nothing retrieved to constrain the response, a model falls back on statistical association, which is serviceable for a well-documented organisation and unreliable for one it holds little about.
02 Related terms
The terms next to this one.
Technical
Retrieval
A model fetching live pages while composing an answer.
Core
Hallucination
A confident model statement that no source supports.
Core
AI Citation
When an assistant attributes part of its answer to your page.
Technical
Vector Embedding
Text turned into numbers, so meaning can be compared by distance.
The work behind it
Tell the models what you are.
Markup that tells models what your business is, on pages that render without JavaScript.