That's interesting because normally with closed backs it is the bass that varies wildly
They use the inward facing microphone to continuously monitor the in situ response and adjust it, up to a point, to deliver the same response across heads and test rigs, regardless of how they actually couple to such heads and rigs. For ANC headphones with a feedback system, it's a byproduct of the feedback system to begin with anyway (although some are a lot better at it than others), but AirPods are also capable of an "ANC-less" "feedback" when ANC is turned off (and I really mean "off", with AirPods and some ANC headphones the feedback system is still active in transparency mode), but it works differently (it needs to monitor a broad signal, such as noise or music, over a short period of time, to calculate the filters).
The APM's variation above 1kHz between measurements is actually quite reasonable for a pair of ANC over-ears, it's a lot, lot worse for headphones such as the Bose 700 or N700NCM2. In the case of the APM I believe that sample variation is low, so it's probably down to coupling, test fixtures' differences, and variation in operators' practices. How they'll behave on your head is, to some degree, unknown in that region, so take measurements of the APM over 800-1k Hz with a rather large pinch of salt. Between 50-800Hz however, you can be very confident that what you'll get on your head is similar to what's been measured.
