Grounding is the practice of basing an AI model's answer on specific external sources, such as retrieved web pages, rather than on the model's internal memory alone. A grounded answer is tied to evidence the system can point to, which often appears to the user as citations.
Grounding reduces invented or outdated information and is what allows AI assistants to cite live sources. It is closely related to retrieval-augmented generation, the mechanism that supplies the sources an answer is grounded in.