Taking feedback from the output eliminates all errors contributed by the regulator. If you are designing for low noise or low drift, then this is the way to do it.
However like Kleinstein says, taking feedback from the output places the regulator within the control loop complicating frequency compensation. Personally I have never had problems doing this but the frequency compensation should be verified empirically which only requires a pulse or function generator and oscilloscope. If you do the frequency compensation this way, then it is possible to work back and derive the AC characteristics of the integrated regulator.