It started out as gate-length and then when we started building 3D transistors with FinFETs and gate-all-around, where the 2-dimensional gate-length is not comparable to "flat" transistors, they had to instead estimate the effective equivalent 2D gate-length that would give the same transistor density.

So the process name is now no longer a measure of any tangible feature size but more a descriptor of transistor density that is loosely consistent with the prior convention.

It seems like 1994 was where process nodes started to not be so correlated with their actual size, according to this IEEE article. In 1994, transistor features were actually smaller than what was advertised, up until the early 2000's, where the naming became smaller than physical size. From what I understand, most of the gains in computing power have come from other improvements in processes and transistor geometry.

I guess the industry never really bothered changing their naming schemes, or couldn't figure out a better way?