Two assistants answer the same question about a coin. One writes three confident paragraphs. The other writes two, names the four readings it used and when each was taken, and flags that one source failed. The second is more useful even when the first is right, because only the second can be audited.
What grounding actually requires
Three things, none of them exotic. The model has to have called something real. The result has to have arrived in a form it can quote. And the system has to require attribution, because a model will happily summarise without saying where anything came from unless told otherwise.
The third is the one most implementations skip, and it is the cheapest.
It changes what the model will assert
This is the part that surprises people. A model instructed to attribute every figure is measurably less inclined to produce figures it cannot attribute. The requirement acts as a filter during generation, not just as formatting afterwards. Removing the place where an invented number could go removes a lot of invented numbers.
Sources are not all equal
A grounded answer should distinguish between what it computed, what it retrieved from a third party and what it found in public reporting. Those have different reliability and different ages, and blending them into one voice destroys the distinction exactly when it matters.
A reading taken a minute ago and a news story from this morning can both be relevant. Presenting them as equally current is a small dishonesty that compounds.
And failures should survive into the answer
If a source was unreachable, the answer should say so. An answer that silently drops a failed input looks complete and is not, and the reader has no way to know which of the two they are holding.