This article investigates the effects of the nonlinear behavior of soils on site response, through various earthquake recordings from the KiK-net database in Japan. This network is composed of more than 688 surface–borehole instruments, from which a characterization of the shear- and compressive-wave velocity profiles down to the borehole depth is available. We selected events with a peak ground acceleration (PGA) at the downhole station of <10 cm=s2 in order to characterize the linear soil behavior by computing the surface to downhole spectral ratios at each site. Modifications of site-response curves computed with strong events (PGA >50 cm=s2) compared to the linear characterization are supposed to be caused by nonlinear soil behavior. To describe the effects of soil nonlinear behavior on site response per event, we propose the percentage of modification (either amplification or deamplification) of the site-response curve compared to the linear evaluation (PNLev, percentage of nonlinearity) and the associated shift frequency (Shev). These parameters are used to estimate the probability that nonlinear site response is significantly different that the linear counterpart. We find that, regardless the site, this probability is important even for low input-motion PGA (values equal to or larger than 30 cm=s2 at downhole sensor). This indicates that nonlinear soil behavior must be taken into account in site-response evaluation for moderate to strong motion. In addition, for 54 sites of the KiK-net database that have recorded at least two strong events (PGA at the downhole station >50 cm=s2), we define four additional parameters that characterize the effects of soil nonlinear behavior on site responses for each site: (1) a PGA threshold (PGAth), defined as the PGA value for which PNLev is higher than 10%; (2) a site-specific PNL for a PGA of 50 cm=s2 (PNLsite); (3) a site-specific shift of the predominant frequency for a PGA of 50 cm=s2 (Shsite); and (4) a frequency from which we observe deamplification between nonlinear and linear site responses (fNL). We observe that nonlinear soil behavior can increase the amplification at frequencies below fNL.We find that fNL lies in between the fundamental and the predominant resonance frequencies of the site response and that sites having VS contrast close to the surface trigger nonlinear behavior at a lower input-motion PGA threshold. These results suggest that nonlinear behavior occurs mostly in the superficial soil layers. Furthermore, by investigating the nonlinear behavior of soils on earthquake horizontal-to-vertical spectral ratios at the surface, we find that they can give satisfactory results (equivalent to the analysis of borehole site responses) for the evaluation of the fNL frequency and shift frequency (Shsite), which indicates that part of the results obtained in this study can be extended to other databases without downhole sensors.