We propose a method to assist the user in the interpretation of the best Bayesian network model indu- ced from data. The method consists in extracting relevant features from the model (e.g. edges, directed paths and Markov blankets) and, then, assessing the con¯dence in them by studying multiple locally optimal models of the data. We prove that our approach to con¯- dence estimation is asymptotically optimal under the faithfulness as- sumption. Experiments with syn- thetic and real data show that the method is accurate and informative.