Live web-search grounding uses native search tools to inject real-time web results and grounding chunks into model context, overriding pure training recall with fresh search information.