To me the magic line with CPU performance is 30%. I think a 30% improvement can be easily felt in day-to-day use. Lower than that and you either have to be really sensitive, have a very demanding use case, or you have to be a benchmark junkie. I almost always upgraded when 30% was on the table.
The IPC from Sandy to Haswell doesn't get you there. Neither do the clock speeds if you have a decent overclocking 2500K/2600K. Adding more cores technically gets you there, but few program utilize 4 cores completely and for the rest you will feel the lower clock per core. All the low hanging fruit has been picked. Now I get that 30% percent back on the energy bill, which considering North American cheap access to power (and power hungry GPUs anyway) isn't an improvement that is worth doing unless you are purposefully trying to move to a new form factor.
To go even further, I can't think of a new TASK that I am doing that my old C2Q desktop couldn't do. Back in the day that 30%+ perfect improvement was enough that maybe you could take on a new task, a new program, or maybe a new game your old computer wouldn't run. Now most games are GPU limited, as they are dragged down by the weak CPUs in console or the fact that the market has stagnated so much. Most of my fun now has been reliving the modern computer revolution on the mobile platform, and strive for the day my phone can do everything that C2Q desktop could do as well.
Actual 21st century compute tasks (like Siri, or Google Now, etc. ) need an army of the very best computers to work- I can't do it on the best desktop possible anyway. Not only are computers "good enough" for normals, it is getting to the point where they are good enough for high-end tasks that can be done on a PC too.
The saving grace for computers is VR. That will need a lot of compute power, and we will prefer it to be local for lag reasons. I hope in five years the reason to upgrade your system will be to run your brand new VR headset. By the time that market gets going maybe ARM will actually be challenging Intel, and the use case will be platform agnostic enough to drive the entire market.
Or not and VR will flop and computer enthusiasm will be a tale we tell our children. Either way it was a fun ride.