Definition
Grounding is connecting a model's answer to retrieved documents so that its claims trace to sources rather than to memory alone. When grounding is on, the answer usually carries citations; when it is off, the model answers from what it absorbed in training.
Why it matters
It decides whether this month's work can matter at all. A newly published page can only influence a grounded answer; an ungrounded answer reflects a corpus that closed long ago.
How to measure it
Look for citations in the answer as the practical signal. Track grounded and ungrounded answers separately if your tooling can distinguish them, because mixing them makes a trend line meaningless.