I believe that total creatine is the most common, especially when using linear combination modeling algorithms like LCModel. The 3 ppm peak should be adequately captured by the Cr and PCr basis functions in most situations.
In all the implementations I’m aware of, the (total) Creatine used for scaling is literally the total raw creatine value that that algorithm would report; there isn’t generally a separate fit just for Creatine referencing (there may be a simplified fit for the initial frequency/phase referencing, but that’s a separate issue).
So if you’re using a basis set approach (LCModel, Osprey, FSL-MRS, QUEST, spant, etc) which yields (unscaled) estimates for Cr and PCr based on the typical metabolite range, say around 0.2-4.2 ppm, then it’ll consider all the features in that range (hence: including the 3.9x peak).
If you’re using a peak-fitting approach (Gannet, AMARES, …) then the situation is a little different; it can be harder to get reliable separation of highly overlapping peaks. I believe Gannet fits a Lorentizan to the 3.0 ppm Cr peak (internally it happens to model this simultaneously with 3.2 Cho, but only area under the estimated Cr peak is used for scaling). AMARES will do basically whatever you ask it to. In that case it depends a lot on your how complete your model is, and how effective the constraints… I expect there’s good argument for using just the clearly resolved 3 ppm peak in this scenario.
They’re heavily overlapped; both Cr and PCr have singlets at 3.0-ish and 3.9-ish ppm.
If you’re using linear combination modeling with a simulated basis set, both of these resonances would be captured during fitting, and you can reference to the tCr directly as reported by the software you use. If you’re using another fitting method (you mentioned AMARES in your edited comment) you could report to the fit of the 3 ppm tCr peak, as long as you compensate for the number of protons contributing to that resonance (3).