On the robustness of primitive words