Grounding

Grounding is the technique of having an AI model pull in outside information, most often a live web search or a document, and build its answer from that instead of purely from what it memorized during training. The idea is that a model's training data has a cutoff date and can be wrong or out of date, while a grounded answer can reflect what's true right now.

Google documents this directly in its Gemini API developer docs as Grounding with Google Search: a feature that lets a Gemini model issue a search and use the results as the basis for its response, with citations back to the sources it used. This is a documented, real product feature, not marketing language.

Grounding is closely related to why an AI system cites sources at all. An answer engine that grounds its response in a live search naturally ends up pointing at the pages it pulled from, which is also how a business becomes a citation. See the citation and answer engine entries for how many sources different engines typically pull in.

Grounding reduces the risk of a model making something up, but it doesn't eliminate it. A grounded model can still misread a source, blend facts from two different sources incorrectly, or state something confidently that the source didn't actually say. See the hallucination entry for what that looks like and how it can go wrong even with retrieval in place.