Brain mapping using magnetic resonance imaging (MRI) is traditionally performed using voxel-wise statistical hypothesis testing. Such mass-univariate approach ignores subtle spatial interactions. The searchlight method, in contrast, uses a multivariate predictive model in each local neighborhood in brain space-named the searchlight. The classification performance is then reported at the center of the searchlight to build an information map. We extend the searchlight technique to take into account additional voxels that can be considered as a meaningful network; i.e., we define a criterion of multivariate connectivity to identify voxels that are statistically dependent on those in searchlight. We coin the term "connectivity searchlight" for the extended searchlight. Using simulated data, we empirically show improved performance for brain regions with low signal-to-noise ratio and recovery of underlying network structures that would otherwise remain hidden. The proposed methodology is general and can be applied to both functional and structural data. We also demonstrate promising results on a well-known fMRI dataset where images of different categories are presented.