Grounding — tying a generated answer to retrieved source documents, so the model reports what it found rather than what it remembers.
Why it decides whether you get named
Ungrounded, a model answers from its weights. Ask it for the best roofer in a mid-sized city and it will reach for whatever was most repeated across the open web — usually a national brand, a directory, or a plausible-sounding invention.
Grounded, it answers from what came back in the search. That is the entire opening for a local business. You are not going to out-remember a national chain. You can comfortably out-document one in your own city.
What makes a page groundable
One claim, stated plainly, in one place, attributable to you, in the language a buyer would actually use. Not a brochure paragraph that takes four sentences to arrive at the point. The system is looking for something it can lift and stand behind.
Grounding is also what separates a citation from a hallucination. A model with nothing to ground on does not return an empty answer — it returns a confident one.
Related: retrieval-augmented generation, hallucination, generative engine optimization.