Optical metasurfaces rely on subwavelength scale nanostructures, which puts significant constraints on nanofabrication accuracies. These constraints are becoming increasingly important, as metasurfaces are maturing toward real applications that require the fabrication of very large area samples. Here, we focus on beam steering gradient metasurfaces and show that perfect nanofabrication does not necessarily equate with best performances: metasurfaces with missing elements can actually be more efficient than intact metasurfaces. Both plasmonic metasurfaces in reflection and dielectric metasurfaces in transmission are investigated. These findings are substantiated by experiments on purposely misfabricated metasurfaces and full-wave calculations. A very efficient quasi-analytical model is also introduced for the design and simulations of metasurfaces; it agrees very well with full wave calculations. Our findings indicate that the substrate properties play a key role in the robustness of a metasurface and the smoothness of the approximated phase gradient controls the device efficiency.