Résumé : Full tensor gravity (FTG) devices provide up to five independent components of the gravity gradient tensor. However, we do not yet have a quantitative understanding of which tensor components or combinations of components are more important to recover a subsurface density model by gravity inversion. This is mainly because different components may be more appropriate in different scenarios or purposes. Knowledge of these components in different environments can aid with selection of optimal selection of component combinations. In this work, we propose to apply stochastic inversion to assess the uncertainty of gravity gradient tensor components and their combinations. The method is therefore a quantitative approach. The applied method here is based on the geostatistical inversion (Gaussian process regression) concept using cokriging. The cokriging variances (variance function of the GP) are found to be a useful indicator for distinguishing the gravity gradient tensor components. This approach is applied to the New Found dataset to demonstrate its effectiveness in real-world applications.