Bayesian, classical, and extended maximum likelihood approaches to estimation of upper limits in experiments with small numbers of signal events are surveyed. The discussion covers only experiments whose outcomes are well described by a Poisson statistic. A new approach, based on the statistical significance of a signal rather than on the number of events in the signal region, is proposed. A toy model and an example of a recent search for the lepton number violating decay τ → µγ are used to illustrate application of the discussed techniques.