Updating a monolithic LLM seems like a harder problem at the moment, and probably would fall under the umbrella of "continual learning", though that sub-field would have their own subdivisions of different methods.
As for how they would potentially perform, those results would be mostly empirically measured, but neither approach would completely remove the possibility of (1) missing results -- which is probably acceptable given Google doesn't always return everything relevant, and (2) hallucinating non-factual responses -- which can be more dangerous, if not from directly instigating / causing harm, then from the incessant worry that it might.