All-occurrence recording is generally thought to yield unbiased results. However, its accuracy and power to detect real effects may depend on the temporal patterning of the observation periods. Although several studies evaluated the suitability of time-sampling methods (one-zero and instantaneous sampling), there has not been an evaluation of various temporal patterns of the ‘all-occurrence’ method to date. Here we present an empirical comparison of time-sampling and all-occurrence recording methods, based on a large data set, with special reference to various temporal patterns of all-occurrence recording. The three different sampling methods were imposed on raw data protocols. The scores, obtained by the use of each method and temporal pattern, were compared with true values from the corresponding continuous video recordings. The accuracy and power of each method was estimated by comparison of medians and magnitudes of effect (rm), respectively. We were able to confirm results of earlier studies on the suitability of time-sampling methods (one-zero and instantaneous sampling), but we found that different temporal patterns of all-occurrence sampling yielded magnitudes of effect that differed from true value magnitudes. We conclude that behavioural studies may reveal different results, even when all conditions are the same, only because of power differences of different temporal pattern of the observation periods, for example 1 × 15 min versus 15 × 1 min. Recommendations are made about appropriate application of methods.