In this paper, we develop data-driven model reduction methods for monotone nonlinear control systems based on a nonlinear version of the dc gain. The nonlinear dc gain is a function of the amplitude of the input and can be used to evaluate the importance of each state variable. In fact, the nonlinear dc gain is directly related to the infinity-induced norm of the system as well as a notion of output reachability. Given the dc gain, model reduction is performed by either truncating not-so-important state variables or aggregating state variables having similar importance. Under such truncation and clustering, monotonicity and boundedness of the nonlinear dc gain are preserved; moreover, these two operations can be approximately performed based on simulation or experimental data alone. This empirical model reduction approach is illustrated by an example of a gene regulatory network.