Hi Yiyi,
Thanks for the question. Broadly speaking the (mathematical) model has two tasks regarding the scale factor. The first is to find a global scale factor that makes all the curves bigger or smaller. If we did absolute normalization of data without errors, this scale factor would always be 1, but it rarely happens so we let the model vary that. If you have a two-phase structural model the mathematical model's other job is to find a relative scale factor for the two phases. We sometimes call this a mixing fraction. If our two components are A and B we might write something like G_tot = f*G_A + (1-f)*G_B.
Now, there is more than one way to set things up in PDFgui. They are mathematically equivalent so it doesn't matter how you do it. However, if you want to follow the logic above, the way that will give the most intuitive results would be to set a global scale factor (which is an attribute of the data, so select the dataset and make a variable @27 or whatever in the scale-factor field there) and then for the scale factors in each of the structural phases give one phase, say A, its own variable (say @50) and the other phase, B, give it (1-@50) in the scale factor position. Your @50 variable will then give you the "proportion" of thing A directly.
Be careful with "proportion", it is strictly the proportion of the signal, which is some kind of weird scattering power weighted thing. If you look in the results section of PDFgui, I think it actually gives the atomic and mass and per unit cell percents of each phase that it extracts from this number.
S