One‐dimensional (1D) linear, equivalent‐linear, and nonlinear site‐response models have been shown to be biased toward underprediction at high frequencies in the aggregate, particularly at small‐to‐medium strains. Because this bias persists among various constitutive models, we hypothesize that breakdowns in the 1D site‐response assumptions and/or poorly characterized soil properties are responsible for the consistent underpredictions. We test four physical hypotheses for this persistent bias using 398 ground motions at 10 selected sites in Japan’s Kiban–Kyoshin (KiK‐net) database that are adequately modeled by 1D wave propagation. Specifically, we (1) apply a depth‐dependent shear‐wave velocity () gradient within layers, (2) decrease the small‐strain damping ratio by half, (3) increase the small‐strain shear modulus by 10%, and (4) randomize the profile. We find that the application of a depth‐dependent gradient and the reduction of the small‐strain damping ratio most greatly reduce the high‐frequency bias; that the randomized profiles sometimes improve predictions at the fundamental site frequency but often lead to greater underpredictions at high frequencies; and that the 10% adjustment of the small‐strain shear modulus has a minimal effect. A significant finding of this study is that overly coarse profiles, which are inadequately sampled in depth, induce considerable underprediction bias in site‐response models at high frequencies. With regard to 1D site‐response model improvement, this study suggests that greater attention should be paid to the coarseness of profiles and excessive impedance contrasts, and that profile corrections using depth‐dependent gradients may be warranted in some cases.