Careful survey design is also a really important element of this. While a dense network of GCPs will allow you distort your 3D model back into shape, it is better if you can minimize the camera model distortion in the first place.
Because of the progressive nature of the bundle adjustment adding photos one/two/few at a time, you can get build up of systematic errors across the model.
The most typical case for this is actually planar acquisition - if you take all of your photos at nadir (straight down) then there is a tradeoff between the radial distortion and the focal length for the lens. The net result of this is a "bowling" of the entire model.
You can either fix this after the fact, with lots of ground control, to restore the model to "flat" (I assume this is what the optimize button does in photoscan), or you can minimize it before hand by careful survey design - though ideally you would do both! You need to have a number of off-nadir images included in the dataset - this will significantly reduce the lens miscalibration.
As mentioned above, you will need some precisely located ground control (>=3) in order to get an accurate tip/tilt for your model. Standard GPS (as on the phantom) will give you a reasonable scale for your model if your survey area is large enough (typically ~3m horizontally, which will be 1% over 600m).
As ARF said though, whichever way you do it, you should definitely use plenty of independent checks the first time you do it.