Accelerometer-derived neutral mass density (NMD) is an important measurement of the variability in upper atmosphere and one of the widely used measurements to calibrate and validate models used for satellite orbit determination and prediction. Providing precise estimates of the true uncertainty of these NMD products is a challenging task but essential for the space weather and geodetic communities. Using multiple data assimilation (DA) experiments and robust statistical techniques, we investigate the uncertainty distribution of three different accelerometer-derived NMD products from the CHAMP satellite mission. Here, in three different DA experiments, we use an ensemble Kalman filter to drive a physics-based model with CHAMP in-situ electron density and temperature data as well as neutral wind estimates from an empirical model. Using a multi-model ensemble comprised of both physical and empirical models, we characterize the error variances among the different NMD products. Our results indicate considerable differences among the CHAMP data sets and also show a pronounced latitudinal dependency for the estimated error distributions. On average, the error estimates for NMD vary in the range 6.5–15.6% of the signal. Our experiments demonstrate that DA considerably enhances the capability of the physical model. We note that the generic strategies applied here may be useful and applicable to other space missions spanning over longer time periods.