Traditional service composition focuses on how to construct optimal composite solutions that satisfy end-to-end QoS constraints raised by users. Although there have been rich well-recognized service composition approaches, most of them ignore an obvious phenomenon: users have to go through a cognition process on unacquainted services to gradually get familiar with these services, and directly using unacquainted services in composite solutions might lead to potential risk on user experiences. QoS fluctuations of services might affect the cognition process to some extent, too. In this paper, we define a new concept anchoring cost to measure such degree of unfamiliarity and give its measurement in terms of historical service usage records and QoS fluctuations. A multi-objective optimization model considering QoS fluctuations and anchoring cost is introduced: besides pursuing the optimality of QoS attributes, to reduce the risk that is caused by importing services with higher anchoring cost into composite solutions is another optimization objective. A set of Pareto-optimal non-dominated composite solutions are obtained by the NSGA-II algorithm and different types of users in terms of risk aversion (e.g., balanced, conservative, risk-taking) may choose their preferred solutions from them. Experiments are conducted on real QoS datasets in three scenarios to demonstrate the rationality and significance of the proposed approach. To the best of our knowledge, this is the first time that cognition cost is incorporated into the service composition problem.