The detection of the optical transient AT2017gfo proved that binary neutron star mergers are progenitors of kilonovae (KNe). Using a combination of numerical-relativity and radiative-transfer simulations, the community has developed sophisticated models for these transients for a wide portion of the expected parameter space. Using these simulations and surrogate models made from them, it has been possible to perform Bayesian inference of the observed signals to infer properties of the ejected matter. It has been pointed out that combining inclination constraints derived from the KN with gravitational-wave measurements increases the accuracy with which binary parameters can be estimated, in particular breaking the distance-inclination degeneracy from gravitational wave inference. To avoid bias from the unknown ejecta geometry, constraints on the inclination angle for AT2017gfo should be insensitive to the employed models. In this work, we compare different assumptions about the ejecta and radiative reprocesses used by the community and we investigate their impact on the parameter inference. While most inferred parameters agree, we find disagreement between posteriors for the inclination angle for different geometries that have been used in the current literature. According to our study, the inclusion of reprocessing of the photons between different ejecta types improves the modeling fits to AT2017gfo and, in some cases, affects the inferred constraints. Our study motivates the inclusion of large similar to 1-mag uncertainties in the KN models employed for Bayesian analysis to capture yet unknown systematics, especially when inferring inclination angles, although smaller uncertainties seem appropriate to capture model systematics for other intrinsic parameters. We can use this method to impose soft constraints on the ejecta geometry of the KN AT2017gfo.