1) Rotation speed [though zero is even better...]
2) Arm positioning time
3) Capacity per platter
Buffer capacity (it's not a cache and cannot work as a cache, as Windows uses hundreds of MB of so-called system cache in the Ram for that use, hence it will never reask a data just read nor written) has little importance. Only if writing (or reading) a huge number of files in parallel, so that Ncq has time to reoptimize the sequence.
But for manufacturers and distributors, cache size is something customers can easily compare... And small Ram chips also get difficult to find. 32MB chips made 256MB modules.
I own two ST3500418AS, their platter is the fastest among 7200rpm (134MB/s measured) BUT their arm is 2ms slower than Hitachi (and maybe Maxtor).
I haven't bought Maxtor nor WD for years for having seen them die.
For a faster arm, I'd take a Hitachi with 320GB platters: the 7k1000b - though I haven't measured it by myself; I have 7k160 which are very agile with similar access times. Take capacities of 160GB, 320GB, 640GB or 1000GB there but not 750GB (3 platters now) nor 500GB.
Head positioning time is all-important for an OS disk. Consider the ST3500418AS like a Hitachi with AAM on: more silent, less agile. Perfect for pictures, videos, downloads - or if you have the OS on an SSD.
Measurement programmes sensible for an OS disk are IOMeter, CrystalDiskMark and to some extent Atto - but not HdTach and others.