In kite power systems, substantial input delay between the actuator and the tethered kite can severely hinder the performance of the control algorithm, limiting the capability of the system to track power-optimal loops. We propose a method that deals with this impediment by using a data-based adaptive filter that predicts future states despite variations in wind conditions, other exogenous disturbances and model mismatch. Moreover, we exploit the geometry of the path on a hemisphere to enhance the guidance algorithm for such kites at a fixed length tether. The objective is to improve the automatic crosswind operation of an airborne wind energy system. To test this under realistic conditions, a small-scale prototype was employed for a series of experiments. The robustness to disturbances and the performance of the algorithm in path following was evaluated for a number of different paths. © 2015 IEEE.