We here turn the general and theoretical question of the complementarity of EEG and MEG for source reconstruction, into a practical empirical one. Precisely, we address the challenge of evaluating multimodal data fusion on real data. For this purpose, we build on the flexibility of Parametric Empirical Bayes, namely for EEG-MEG data fusion, group level inference and formal hypothesis testing. The proposed approach follows a two-step procedure by first using unimodal or multimodal inference to derive a cortical solution at the group level; and second by using this solution as a prior model for single subject level inference based on either unimodal or multimodal data. Interestingly, for inference based on the same data (EEG, MEG or both), one can then formally compare, as alternative hypotheses, the relative plausibility of the two unimodal and the multimodal group priors. Using auditory data, we show that this approach enables to draw important conclusions, namely on (i) the superiority of multimodal inference, (ii) the greater spatial sensitivity of MEG compared to EEG, (iii) the ability of EEG data alone to source reconstruct temporal lobe activity, (iv) the usefulness of EEG to improve MEG based source reconstruction. Importantly, we largely reproduce those findings over two different experimental conditions. We here focused on Mismatch Negativity (MMN) responses for which generators have been extensively investigated with little homogeneity in the reported results. Our multimodal inference at the group level revealed spatio-temporal activity within the supratemporal plane with a precision which, to our knowledge, has never been achieved before with non-invasive recordings.