The mass-loss rate of donor stars in cataclysmic variables (CVs) is of paramount importance in the evolution of short-period CVs. Observed donors are oversized in comparison with those of isolated single stars of the same mass, which is thought to be a consequence of the mass loss. Using the empirical mass-radius relation of CVs and the homologous approximation for changes in effective temperature T_2, orbital period P, and luminosity of the donor with the stellar radius, we find the semi-empirical mass-loss rate M2_dot of CVs as a function of P. The derived M2_dot is at ~10^(-9.5)-10^(-10) Msun/yr and depends weakly on P when P > 90 min, while it declines very rapidly towards the minimum period when P < 90 min, emulating the P-T_2 relation. Due to strong deviation from thermal equilibrium caused by the mass loss, the semi-empirical M2_dot is significantly different from, and has a less-pronounced turnaround behavior with P than suggested by previous numerical models. The semi-empirical P-M2_dot relation is consistent with the angular momentum loss due to gravitational wave emission, and strongly suggests that CV secondaries with 0.075 Msun < M_2 < 0.2 Msun are less than 2 Gyrs old. When applied to selected eclipsing CVs, our semi-empirical mass-loss rates are in good agreement with the accretion rates derived from the effective temperatures T_1 of white dwarfs, suggesting that M2_dot can be used to reliably infer T_2 from T_1. Based on the semi-empirical M2_dot, SDSS 1501 and 1433 systems that were previously identified as post-bounce CVs have yet to reach the minimal period.