Recordings from two dense arrays deployed at Argostoli, Cephalonia Island, Greece, are analyzed with three objectives: (1) exploring to what extent the diffracted surface waves influence the seismic wavefield at a rock site, (2) investigating the loss of coherency of ground motions, and (3) comparing the results for two nearby sites with different soil conditions. The two dense arrays under consideration consist of 21 velocimeters encompassing a central station in four concentric circles with diameters 10–180 m at the soft‐soil site and 20–360 m at the rock site. The datasets include 40 or more events with relatively homogeneous distributions in epicentral distances (10–200 km), magnitudes (2–5), and back azimuths. The wavefields are analyzed using the MUSIQUE algorithm: back azimuth and slowness of dominant incoming waves are extracted, and Love and Rayleigh waves are identified. Average lagged coherency estimates are provided for interstation distances 10–20, 20–30, 30–40, and 80–90 m. Coherency is observed to be generally larger on the rock site compared with the soft‐soil site, especially at frequencies below 5 Hz. At the soil site, lower coherency is observed for pairs along the valley‐perpendicular direction while no such directional dependence is observed at the rock. Although about 40%–60% of the seismogram energy at the soft‐soil site could be associated with diffracted surface waves (Love and Rayleigh) propagating along the valley‐perpendicular direction, only about 20% of energy at the rock site could be characterized as diffracted surface waves. Comparison with the widely quoted parametric models reveals little correlation with the observed decay of coherency at both sites. These significant differences between rock and soft‐soil array results indicate that the spatial incoherency is largely site dependent and is likely to be closely related to the formation of locally generated wavefield.