Why it is the hardest step
Answers drift on their own: the same question gets different answers from the same model on different days. Without a baseline measured before the change and a single changed factor, a difference after the change could be noise, a model update or a change somewhere else on the web.
What does not count
A crawl after publishing, a citation after publishing, a better answer after publishing. Each is a sequence, not a cause. The rule is published at /evidence.