Thank you for writing that post, it's quite the amazing insight you got there.
I agree that using NBD_LW instead of NBD_ON could give interesting results, and could certainly provide a more realistic assessment of speakers that show problems in the response directly in front of the tweeter, such as coaxials (IN-8, LS50). I have two reservations, however.
First, in the absence of any other information, most people will assume they will get the best sound by pointing the speaker directly at them. They might not do that all the time (that's often an unrealistic use case), but they might do that when they want to do some critical listening. In such cases, they are in fact getting the on-axis response.
Second, the problem with the listening window is that's it's an average. You can't actually get that response no matter what angle you listen at, because it doesn't correspond to any particular angle. For this reason the score might not match what a typical listener would rate the speaker, even while listening slightly off-axis - it is likely an overestimate from excessive smoothing. One could imagine pathological cases where a speaker has a huge frequency response abnormality at some angles, and the opposite abnormality at other angles, which would average out to "perfect". Though maybe I'm being overparanoid here.
I can't help but think that these problems would be avoided if manufacturers carefully specified the optimal listening axis. In the case of a coaxial, for example, the manufacturer could indicate in the manual that the best sound is obtained by listening slightly off-axis (e.g. "no toe-in"). Then if
@amirm measures the speaker using such an offset reference axis, the results will be good even with the original formula.
Oh, and let's not forget
@MZKM's "
it won't end there" prediction - this is on track to prove him right. At some point we're gonna have to agree upon a small set of dimensions by which to break down this data, otherwise Sheets will explode under the sheer number of tabs