This study proposes a new LLM-as-a-judge method for relative evaluation of LLM responses based on both required criteria (i.e., minimal conditions that responses must satisfy) and bonus criteria (i.e., desirable additional qualities). Recent LLM-as-a-judge methods have improved the stability and accuracy of relative evaluation through the use of automatically generated criteria. However, existing methods rely solely on the required criteria, which may limit their ability to distinguish closely matched responses. To address this issue, the proposed method first evaluates responses based on required criteria and, in the case of a tie, further evaluates them using bonus criteria, enabling more fine-grained comparisons. Experiments on the Writing, Roleplay, and Humanities categories of the MT-Bench dataset demonstrate that the proposed method achieves higher agreement with human judgments than existing methods.