Damage localization based on ambient vibration data in combination with finite element models can be challenging, in particular due to the large number of parameters in the model and noisy measurement data. Changes in different structural parameters can cause similar changes in datadriven features, and vice versa, it can be challenging to identify which parameter caused the deviation in the data. The problem is ill-conditioned and slight variations in the features, due to inherent statistical uncertainty, can lead to significant errors in the result interpretation. A possible solution is sensitivity-based statistical tests in combination with a parameter clustering approach that considers the uncertainties of data-driven features. In this context, this paper introduces the concept of damage localizability, and provides a framework to evaluate it based on the minimum detectable parameter changes, possible false alarms in unchanged parameters, as well as the achievable damage localization resolution. Since clustering approaches depend on user-defined hyperparameters, such as the number of clusters, the second objective of this paper is to optimize the performance of the damage localization, by adjusting the hyperparameters for clustering. A particular strength of the approach is that the analysis can be conducted based on data and a numerical model from the undamaged structure alone, making it a suitable approach to assess and to optimize the diagnosis performance before damage occurs. For proof of concept, a laboratory case study on a simply-supported steel beam is presented, where the localizability of mass changes is analyzed and optimized.