A posteriori error estimators, gradient recovery by averaging, and superconvergence