Re: How to count: Let us count the ways!
It's definitely mushy, though I think we can partially address some of these questions. I honestly hadn't thought about how demonstration models sitting in, say, Best Buy or an Apple store might affect these statistics, but I suppose that they could. For Net Applications, if someone playing with a display computer happens to access a website in the relevant network, I believe that computer would count as a single "user." If another shopper at the store comes up to the display computer a few minutes later and happens to load another site in the Net Applications network, the second shopper wouldn't count as an additional user, and thus would not further inflate that machine's statistical impact. StatCounter, on the other hand, would count both shoppers toward an aggregate usage statistic, however.
As for PCs sitting in warehouses and so on... I don't think those have any influence. To count towards either sets of data, a PC has to access the public Internet and, moreover, load a page that one of the web-tracking agencies happens to have in its network. Likewise, if some businesses have XP machines running isolated apps on closed networks, I don't see how Net Applications or StatCounter would see them.
Neither methodology is perfect, but Net Applications tries to count how many people are using each OS during a given month-- an interesting if intrinsically limited statistic, given that not all users are equally interesting, depending on what you want to know. StatCounter tries to measure which OSes are contributing the most web traffic in a given month-- a different, also interesting, but also intrinsically limited statistic. A population of extreme power users could heavily sway the StatCounter statistics whereas a couple million aging computer turned on once a month for email might inflate an older OS's importance in the Net Applications numbers-- and those are best case scenarios, assuming the numbers aren't corrupted in other ways (which they almost certainly are). We don't have particular insight into the specific sites scanned, for example-- how representative they are for this sort of macro-scale extrapolation, and so on. So it's an imprecise picture, to be sure-- but if we take it with those limitations, I think some useful trends can be discerned—e.g. all those cheap Windows 8.1 devices have most likely driven up sales.