Glossary/Grounding
Grounding
Definition
Grounding is when a language model bases its answer on retrieved external sources and citations rather than on its own trained-in memory.
A model has two ways to answer: from its parametric memory, the patterns learned during training, or from documents fetched at the moment of the question. Grounding means the answer is tied to those fetched sources, which the model can then cite.
Grounded answers are generally more current and more verifiable, because the claims trace back to specific retrieved pages. This is why engines that ground their answers, like Perplexity and Google AI Overviews, tend to show links to their sources.
Grounding is the mechanism that makes optimization possible. If a model only answered from memory, there would be little a firm could do in the moment; because grounded engines retrieve live sources, being a retrievable, trustworthy source directly affects what they say.