Extracting the area where people have difficulty evacuating (hereafter difficult-to-evacuate areas, DEA) when tsunamis hit after an earthquake is important for effective disaster mitigation measures. The DEA was conventionally extracted by simply considering the walking speed, distance to the evacuation destination, and time needed for evacuation after considering the estimated tsunami inundation area. However, evaluating the DEA from such a simple scheme is insufficient because the behavior of residents and the road conditions to the evacuation destinations after an earthquake are not properly reflected in the scheme. In this study, agent-based tsunami evacuation simulations that can reflect the behavior of residents and real-time changes in the situation were conducted in Zihuatanejo, Guerrero, Mexico. It is a prime sightseeing destination under the high risk of megathrust events in the Guerrero Gap. First, by checking the simulation images at the tsunami arrival time, bottleneck locations were identified, and five additional models with different measures for the bottleneck locations were constructed and tested to find the best model with 195 casualties. Then, focusing on the best model, three indices for the casualties were proposed to extract the DEA effectively and quantitatively, and numerical analyses using the three indices was conducted. Finally, the subdistrict in the center of the target area (subdistrict 5) was quantitatively found to be the district that should be given the highest priority for measures. Moreover, an example model with a new measure in subdistrict 5 was validated to have 101 casualties. The key points for applying the proposed method for extraction of DEA in other areas are summarized.