This paper analyzes several systematic errors affecting sea surface gradients derived from Seasat, Geosat/ERM, Geosat/GM, ERS-1/35d, ERS-1/GM and TOPEX/POSEIDON altimetry. Considering the data noises, the conclusion is: (1) only Seasat needs to correct for the non-geocentricity induced error, (2) only Seasat and Geosat/GM need to correct for the one cycle per revolution error, (3) only Seasat, ERS-1/GM and Geosat/GM need to correct for the tide model error; over shallow waters it is suggested to use a local tide model not solely from altimetry. The effects of the sea surface topography on gravity and geoid computations from altimetry are significant over areas with major oceanographic phenomena. In conclusion, sea surface gradient is a better data type than sea surface height. Sea surface gradients from altimetry, land gravity anomalies, ship gravity anomalies and elevation data were then used to calculate the geoid over Taiwan by least-squares collocation. The inclusion of sea surface gradients improves the geoid prediction by 27% when comparing the GPS-derived and the predicted geoidal heights, and by 30% when comparing the observed and the geoid-derived deflections of the vertical. The predicted geoid along coastal areas is accurate to 2 cm and can help GPS to do the third-order leveling.