Open and dynamic environments lead to inherent uncertainty of Web service QoS (Quality of Service). Thus, the Web service composition which consists of these services will be necessarily random of the QoS. The requirements of the QoS of Web service composition may not be certainly satisfied. We use a simulation approach named Importance Sampling to analyze the QoS probability of Web service composition in stochastic PERT network. In this paper, we propose a relatively simple distribution function and introduce a weighting function to ensure that the estimating of the target distribution function is an unbiased estimation. We conduct a comparison of Importance Sampling technique with Monte Carlo simulation about the rationality and the operation efficiency based on the actual QoS data of Web services by experiment. The experimental results prove that the Importance Sampling technique has better precision and higher efficiency than Monte Carlo simulation.