Grounding is tying an AI model’s answer to specific source material so its claims can be traced and checked.
A grounded answer cites what it relied on; an ungrounded one comes purely from the model’s training. Google, for example, offers "grounding with Google Search" as a feature for developers building on Gemini.
Grounding reduces, but does not remove, made up statements. A model can still misread a source or merge two of them.
Content that is easy to ground against tends to share a few traits: one claim per sentence, the publisher and date beside each figure, and no hedging that leaves the reader unsure what is actually being asserted.