We test a unified observation model for estimating surface-loading-induced geocenter motion using GPS. In principle, this model is more complete than current methods, since both the translation and deformation of the network are modeled in a frame at the center of mass of the entire Earth system. Real and synthetic data for six different GPS analyses over the period 1997.25--2004.25 are used to (1) build a comprehensive appraisal of the errors and (2) compare this unified approach with the alternatives. The network shift approach is found to perform particularly poorly with GPS. Furthermore, erroneously estimating additional scale changes with this approach can suggest an apparently significant seasonal variation which is due to real loading. An alternative to the network shift approach involves modeling degree-1 and possibly higher-degree deformations of the solid Earth in a realization of the center of figure frame. This approach is shown to be more robust for unevenly distributed networks. We find that a unified approach gives the lowest formal error of geocenter motion, smaller differences from the true value when using synthetic data, the best agreement between five different GPS analyses, and the closest (submillimeter) agreement with the geocenter motion predicted from loading models and estimated using satellite laser ranging. For five different GPS analyses, best estimates of annual geocenter motion have a weighted root-mean-square agreement of 0.6, 0.6, and 0.8 mm in amplitude and 21¿, 22¿, and 22¿ in phase for x, y, and z, respectively.