And what is real world usage? If you know that Geekbench isn't real world usage then you must know what real world usage is.
In my opinion real world usage (CPU part) varies a lot for different users, but it is always a sum of different "synthetic" math tasks or other simple data manipulation.
I agree completely on the second part.
Personally if I could change geekbench.
Compiler support targeting all relevant extensions for ARM, x86, etc. Less encryption/decompression (Do you really need JPEG and PNG tests? Can't you simply put that under "media compression"). Memory latency tests (just as important as bandwidth). Virtual machine (JS) performance. Character recognition.
The point is if you are going to use "synthetic" math tests then exploit them to the fullest. Geekbench seems to say "We will use a bunch of synthetic math tests to evaluate real world performance but wait, don't make it too synthetic, after all, these completely synthetic math test must somehow apply to the real world".
Now I understand your dilemma, these tests need to run on a variety of platforms so they can't for instance use 4 GB of RAM or something like that.