First, strong‐motion recordings from an Mw 4.4 earthquake with symmetrical rupture were regarded as empirical Green’s functions (EGFs) to simulate ground motions produced by the 2013 Mw 6.6 Lushan mainshock using the two‐step stochastic EGF method proposed by Kohrs‐Sansorny et al. (2005). In this simulation, the source of the Lushan mainshock was set to consist only of an asperity in the vicinity of the hypocenter, which was determined according to the statistical properties of asperity. The good agreement between the simulated ground motions and the observed recordings at short periods (<2.0 s) validates this method. Furthermore, recordings from an Mw 4.5 earthquake with significant rupture directivity were used as EGFs to simulate ground motions of the Lushan mainshock. The rupture directivity mainly exerts significant effects on the simulated short‐period ground motions, which result in large discrepancies with the observed recordings. Simulated ground motions are much higher than the observed ones at directive stations, but lower at antidirective stations. More significant effects are observed at directive stations than antidirective stations. Using the azimuthal apparent corner frequency to consider the rupture directivity in the two‐step stochastic EGF method, the good reproductions of the observed recordings are obtained at short periods, which indicates that the rupture directivity at short periods can be simulated well. Finally, ground motions from a series of Lushan‐like earthquakes assuming various rupture directivities were simulated by adjusting the apparent corner frequency. The simulated ground motions show significant directivity effects. The two‐step stochastic EGF method provides an effective approach to predicting ground motions for future earthquakes because far fewer source parameters are required.