To muddy things a bit, the probe will have internal compensation that can aim to flatten the response and I think that this will make it harder to predict how much droop will actually be seen using this test method.
For lots of reasons I much prefer to look at how the probe performs in the time domain using a pulsed source. However, it might be interesting to see how the probes compare in the frequency domain as well.
I guess the thing to do is try it... I'll put some code together and try testing using this 50R | 50R method as used by Tektronix. I've got a 1GHz scope and a decent (dual 50R) splitter rated to many GHz and various sig gens that are suitable for this. I've also got a decent RF power meter to use as the detector in the feedback arm of the system.
Note that the scope probe loading can't pull down the source voltage at the input port of the splitter. This is because of the feedback in the system. This node will appear as an ultra low impedance voltage source. So the probe can't affect this voltage if the system can remain in closed loop. However, the probe loading can pull down the voltage across the termination resistor. My simulation showed it can pull it down by about 3.5dB by 500MHz.
