Monthly is the right cadence for most businesses. Weekly produces noise that looks like signal, because model updates move results independently of your actions. Quarterly is too slow to connect a change you made to a result you got.
Measure the same prompt set each time. Changing the set mid-programme invalidates every comparison you have.
Take two baseline readings a week apart before starting, so you know what normal variance looks like.
Re-verify the prompt set annually as buyer language shifts.