Geosat sea-level variations are assimilated into a nonlinear shallow-water model of the Indian Ocean. After spinup, the model forced by observed winds over 1985–1988 provides a first guess for initial conditions on November 1986 and a mean thermocline depth used as a reference surface for altimetric observations. Data are assimilated during one year and north of 20°S. The model-data misfit is minimized using the adjoint model with respect to the initial conditions and the reference surface. The assimilation algorithm converges fast and steadily. The monthly rms difference between model and data is reduced from 21 to 16 m. The optimal thermocline surface is significantly affected by assimilation. Large changes occur along the paths of the throughflow and South Equatorial Current. It is found that the optimal mean state is valid only for the period of assimilation. It degrades the agreement between observations and simulations for the following year because of the large interannual event that took place over 1986–1988. With an assumed 7-cm observation error, model and data are overall consistent. But the model is fundamentally inconsistent with data in the Bay of Bengal or in the Indonesian Throughflow region. Examples of erroneous conclusions about model-data consistency are given when data are assimilated with faked assumptions on the observation error structure.