Daily surface incident solar radiation (<em>R</em>_s) from reanalysis products is widely used to assess the reliability, stability, and intermittency of photovoltaic power generation and other applications, owing to their full spatiotemporal coverage and physical consistency. However, whether reanalyses faithfully represent the characteristics of extreme <em>R</em>_s remains insufficiently evaluated. To address this, we assessed the performance of two major reanalysis products, ERA5 and MERRA-2, in simulating mean and extreme <em>R</em>_s across China during 1961&ndash;2016 with the help of precious ground-based observations of daily <em>R</em>_s. Observational analysis of mean <em>R</em>_s shows pronounced solar dimming before 1994, followed by weak brightening. Extremely high <em>R</em>_s (<em>R</em>_s90p) frequency presents multidecadal variations broadly consistent with mean <em>R</em>_s, whereas its magnitude increases persistently from 1980 onward. Conversely, extremely low <em>R</em>_s (<em>R</em>_s10p) frequency increases steadily during 1961&ndash;2016, while its magnitude decreases until 1994 and then reverses thereafter. Although ERA5 broadly captures most segmented variations of both extremes, it underestimates by over 90 % the pre-1980 decline in <em>R</em>_s90p frequency and the post-1994 increases in the frequencies of both extremes. MERRA-2 performs worse than ERA5, particularly in failing to reproduce the evolution of <em>R</em>_s90p frequency and magnitude. The attribution analysis reveals that trend biases in <em>R</em>_s90p are predominantly controlled by those in clear-sky <em>R</em>_s in reanalysis products whereas cloud- and aerosol-related processes jointly influence trend biases in <em>R</em>_s10p. These findings highlight the limitations of current reanalyses in simulating extreme <em>R</em>_s and the need to interpret cautiously their long-term trends in climate and solar energy applications.