Measuring opinion similarity between participants is an important strategy to reduce the chance of making and applying inappropriate decisions in multi-criteria group decision making applications. Due to the small-sized opinion data and the varieties of opinion representations, measuring the similarity between opinions is difficult and has not been well-studied in developing decision support. Considering that the similarity changes with the number of concerned criteria, this paper develops a gradual aggregation algorithm and establishes a three-level-similarity measuring (TLSM) method based on it to measure the opinion similarity at the assessment level, the criterion level and the problem level. Two applications of the TLSM method on social policy selection and energy policy evaluation are conducted. The study indicates that the TLSM method can effectively measure the similarity between opinions in small-size with possibly missing values and simulate the dynamic generation of a decision.