A response informed and supported by external information. Grounding does not independently guarantee 100% factual accuracy.
A grounded answer is informed by information supplied to the model or retrieved for the question. It should reflect that material rather than introduce unsupported details. The label is about the relationship between the answer and its supporting context, not a promise of perfect accuracy.
Two checks are useful. First, does the source actually support what the answer says? Second, is the source itself accurate, current and applicable? An answer can follow an outdated document faithfully, or it can misstate a correct document. Either situation can mislead the reader.
A visible citation can help someone inspect the evidence, but its presence is not enough. Review the specific passage associated with the claim, including qualifications and exceptions. Important conclusions may require additional evidence even when part of the response is well supported.
Takeaway: Use grounded answers as inspectable explanations, not substitutes for verification where errors would matter. Make the supporting information accessible to the appropriate reviewer.
Example
An assistant summarizes a service agreement that covers weekday support only. A grounded response should preserve that condition. If it says support is available every day, adding a link to the agreement does not make the broader claim supported.
What to check
Check the answer at the claim level rather than approving it because one source looks reputable. Distinguish missing evidence, stale evidence and an incorrect interpretation so that the underlying issue can be repaired.
For more detail, see Gao et al.: Enabling Large Language Models to Generate Text with Citations.