Nowadays, web search becomes more and more popular all over the world. Many researchers and developers have done lots of studies on behaviors of search users. In practice, the full understanding of these behaviors can not only help to evaluate the usefulness of newly-developed ranking algorithms and other changes of search engine, but also to guide the growth direction of search engine. As far as we know, most of past work are mainly focused on single search evaluation, which do promote the rapid development of search engine in early stage. However,these page-level behaviors are so limited that can no longer give explicit feedbacks on minor changes of the search engine. We think that it will be more accurate and sensitive when more information on search session are provided. In this paper, a session level evaluation method is proposed. The session-level features are retrieved and carefully analyzed. Some linear and non-linear features which can reflect the final degree of satisfaction are chosen and adopted in evaluation models. A two-layer hybrid evaluation model with different granularity, which can achieve good precision and recall, is designed and trained. Lots of real experiments are evaluated by the model, the result shows it achieved a higher accuracy performance than traditional page-level evaluation metrics. Furthermore, for practical application, it is important to interpret the reason of each session's satisfaction judgement. In all, a session-level evaluation model with improved performance and well capability on interpretation is proposed and applied in real practice in search engine companies.