A primitive equation ocean model has been applied to the study of mesoscale ocean dynamics and prediction using sea surface height information derived from a satellite‐borne altimeter. Results from a model of the Gulf of Mexico were directly compared with the altimeter data from GEOS 3 and Seasat reported by Marsh et al. (1984) and in situ hydrographic data analyzed by Maul and Herman (1985). In the eastern Gulf the amplitude and position of relative maxima of sea surface height variability were found to be similar for model, altimeter, and in situ data. In the central Gulf the model and the altimeter‐derived sea surface height variability maps were similar but differed considerably from the in situ results. The mean sea surface in the model and from the in situ data also agreed well in the eastern Gulf, but there was little agreement in the central Gulf. While there is no geoid adequate for determining an independent altimetric estimate of the mean dynamic height of the sea surface in the Gulf by differencing the geoid with the altimeter‐derived mean sea surface, the model dynamic mean sea surface can be subtracted from the mean surface of Marsh et al. (1984) to obtain a new estimate of the Gulf geoid, which may be the best geoid available in that region for several years. A benchmark experiment (“truth”) was integrated to statistical equilibrium and compared with results from four experiments in which the model was initialized with fields modified from archived benchmark data. The experiments differed only in the initialization fields. Each experiment was integrated for 100 days with inflow transport remaining constant throughout the integration. The idealized experiments were initialized geostrophically (1) with the exact sea surface and pycnocline height fields, (2) with only the exact sea surface heights and a pycnocline assumed to compensate it such that the deep pressure perturbations vanished, (3) just as in case 2 but with a geoid error component added on small spatial scales, and (4) just as in case 3 except the geoid error model included an additional contribution in strong geoid gradient regions. The sequence of four numerical experiments showed that (1) geostrophic initialization with exact sea surface and pycnocline height information provided accurate forecasts to 100 days, (2) even when only the sea surface height information was provided to the numerical model for geostrophic initialization and deep pressure perturbations assumed to vanish at the initial time, the forecasts were superior to persistence or climatology over the 100‐day forecast period, and (3) errors in the geoid on spatial scales comparable to the grid resolution of the model did not seriously degrade the forecast, even in dynamically active regions with large gradients in the geoid height where instability processes were observed to occur.