An ungrounded answer comes from training-data recall, which can be stale or invented. Grounding puts real evidence in the context — search results, your documentation, a database query — and instructs the model to answer from it, ideally with citations. It's the working antidote to hallucination and the reason "chat with your docs" products exist.

Grounding is only as good as its retrieval: fetch the wrong passage and you've grounded the answer in irrelevance. Checking answers against their cited sources remains the honest test.