Because of the new regulatory requirements that hazards have to be estimated in probabilistic terms, the number of probabilistic hazard studies conducted has recently been increasing. The present study aims at defining the possibilities and limits for comparing predictions from these studies and observations. Comparison tests based directly on the rate of ground-motion occurrences are favored over the rate of earthquake occurrences. Based on the properties of Poisson processes, the minimum time window ensuring reliable occurrence rate estimates at a site is computed and evaluated. For example, for ground motions with a 475-yr return period at a site, a minimal 12,000-yr observation time window is required for estimating the rate with a 20% uncertainty (coefficient of variation: standard deviation divided by the mean). These values are not dependent on the seismicity level of the regions under study. An analysis of recorded ground motions at the stations of the permanent French accelerometer network shows that at best, the occurrence rates can be estimated with an accuracy of 30% for very low acceleration levels (0.0001–0.001g for the station STET). The same analysis, carried out at two stations with longer recording histories and located in higher seismicity regions (Greece and California), provides ground-motion levels up to 0.1g. Therefore, the question posed is can the results of a comparison test at low acceleration levels be generalized to higher acceleration levels, even if using a ground-motion prediction equation uniformly valid for a wide range of accelerations?