Grounding means tying an AI's answer to real, checkable sources rather than letting it speak from memory alone. A grounded response does not just tell you the answer — it shows you where the answer came from, so you can verify it. It is the practical opposite of a hallucination.
The most common way to ground a model is retrieval-augmented generation: fetch the relevant documents first, then have the model answer from them and cite them. The distinction worth keeping is that RAG is the technique and grounding is the goal — answers anchored to evidence you can inspect.
Why it matters at your desk. For anyone whose work has consequences, grounding is the feature that turns AI from a risky shortcut into a usable assistant. It is why Perplexity shows links, why Consensus points to the underlying studies, and why Harvey answers from your matter's actual documents — each one lets you trace a claim back to its source instead of trusting the model's word.
What to watch for: grounding raises the floor but does not remove your job. A model can cite a real source and still misread it, or ground an answer in a document that is itself wrong. Citations make verification possible and fast — they do not make it optional.