Quantitative structure property relationship (QSPR) models allow for estimation of unknown substance properties. Their accuracy is usually described by error metrics that compare the result of the estimation to an experimental value. However, it has to be kept in mind that measured values are not perfectly accurate as well. The actual reliability of a QSPR model is therefore often significantly lower than indicated by conventional error metrics. To deal with this issue, the combined experimental and predictive uncertainty (CEPU) is proposed. Analogously to error propagation in evaluation of experimental works, the reliability of estimation methods can be accessed by the root of the sum of the different squared contributions to overall uncertainty. Such a metric is able to describe the effects of experimental errors with uniform algebraic sign as well as of randomized experimental errors.