We discuss the commonly encountered problem when optimizing nuclear magnetic resonance (NMR) pulses using optimal control that the otherwise very precise NMR theory does not provide as excellent agreement with experiments. We hypothesize that this disagreement is due to phase transients in the pulse due to abrupt phase and amplitude changes resulting in a large bandwidth. We apply the gradient optimization using parametrization algorithm that gives high fidelity pulses with a low bandwidth compared to the typical gradient ascent pulse engineering pulses. Our results obtain a better agreement between experiments and simulations supporting our hypothesis and solution to the problem.