In this paper, we discuss the optimal H-2 model order reduction (MOR) problem for bilinear systems. The H-2 optimal MOR problem of bilinear systems is considered as the minimisation problem on Grassmann manifold, which is stored as a quotient space of the noncompact Stifiel manifold. Grassmann manifold whose tangent space is endowed with a Riemannian metric is a Riemannian manifold. In its tangent space equipped with the Riemannian metric, we derive the negative gradients of the cost function, i.e. the steepest descent direction of the cost function. After that, the formulas of geodesic on Grassmann manifold are given. Then we perform linear searches along geodesics to obtain the optimal solutions. Thereby, a two-sided MOR iterative algorithm is proposed to construct an order-reduced bilinear system, which is used to simulate the output and input responses of the original bilinear system. Numerical examples demonstrate the effectiveness of our algorithm.