An approach is developed to build multivariate probabilistic seismic demand models (PSDMs) of multicomponent structures based on the coupling of multiple-stripe analysis and Gaussian mixture models. The proposed methodology is eminently flexible in terms of adopted assumptions, and a classic highway bridge in Eastern Canada is used to present an application of the new approach and to investigate its impact on seismic fragility analysis. Traditional PSDM methods employ lognormal distribution and linear correlation between pairs of components to fit the seismic response data, which may lead to poor statistical modeling. Using ground motion records rigorously selected for the investigated site, data are generated via response history analysis, and appropriate statistical tests are then performed to show that these hypotheses are not always valid on the response data of the case-study bridge. The clustering feature of the proposed methodology allows the construction of a multivariate PSDM with refined fitting to the correlated response data, introducing low bias into the fragility functions and mean annual frequency of violating damage states, which are crucial features for decision making in the context of performance-based seismic engineering.