Accurate determination of earthquake depth is important, but particularly challenging. We develop a new method to determine the earthquake source depth by stacking multiple-station autocorrelograms (SMAC) of seismic data. The basic concept of SMAC method is to enhance the coherent surface reflected energy by autocorrelation and stacking, and uses the surface reflected energy to determine the source depth. Autocorrelation effectively enhances the energy of the seismic phases related to the source depth, while stacking the autocorrelograms of array data further improves the signal-to-noise ratio of the energy. The procedures are applied using both the main SH waves and the coda. Using coda waves, the method extracts the two-way traveltime between the source and the surface by stacking the coda wave autocorrelations over an array of seismic observations. Using the main SH waves, SMAC explores the interference of the down-going Moho reflected SmS wave and up-going surface bounced Moho reflected sSmS wave by autocorrelation. The autocorrelograms are then stacked along theoretical differential sSmS–SmS traveltime predicted for all potential source depths, and the source depth is determined to be the one that produces the maximum stacking energy. Synthetic tests demonstrate the validity of the SMAC procedures. As an example of application, we apply the SMAC method to determine the source depth of an earthquake occurring in Japan Island. The procedures of using both main SH waves and their coda waves yield robust surface reflected energy and a consistent source depth. The error of the depth estimation is less than 1 km assuming an uncertainty of 10 per cent in the averaged crust velocity.