We explore nonlinear site effects in the new Japanese ground‐motion dataset compiled by Bahrampouri et al. (2020). Following the approach of Seyhan and Stewart (2014), we evaluate the decrease of soil amplification according to the increasing and corresponding ground motion on surface rock (). To better predict the rock ground motion associated with each record, we take into account the between‐event variability of the ground motion, and to better evaluate the impact of nonlinearity, we correct observed ground motion on soil by the site‐specific linear amplification. Instead of grouping the stations by site‐response proxy, we focus on individual stations with several strong‐motion records. We develop a framework to test recently published nonlinear site amplification models against a linear site amplification model and compare the results with recent building codes that include nonlinearity. The results show that the site response varies greatly from site to site, indicating that conventional site proxies, such as , are not sufficient to characterize nonlinear site response. Out of all of the Kiban–Kyoshin network stations, 20 stations are selected as having recorded sufficient data to be used in the test. Out of these 20 stations, five stations show signs of nonlinearity, that is, the nonlinear models performed better than the linear‐amplification model for all periods . For most sites, however, the linear site amplification models get the best score. This suggest that, for the range of predicted rock motion considered in this study (peak ground acceleration ), nonlinearity may not have a sufficiently large impact on soil ground motion to justify the use of nonlinear site terms in ground‐motion functional forms and seismic building codes for such moderate‐level shaking.