The fifth generation (5G) mobile communication technologies are attracting a lot of attention in terms of accommodating the huge traffic expected in the future. Millimeter wave communications, which utilize wide frequency bands, are attracting attention for the realization of the high capacity required in the 5G era. In millimeter wave communications, beamforming with massive antennas is expected to play a very important role in compensating the large propagation loss of millimeter waves. Because massive beamforming yields narrow beams, the search for the optimal beam could have considerable impact on the system. In this paper, we propose a new beam search method that can reduce the load of beam search significantly while keeping beamforming gain almost the same as that of the conventional method. The proposed method consists of three stages with the creation of a set of candidate beams in the first stage, selection of an initial beam in the second stage, and refinement of the selected beam in the third stage. In the first stage, the created set of candidate beams contains beams of various widths instead of beams of a uniform width to reduce the number of candidate beams in the set. Here, we leverage the property of millimeter waves according to which the fluctuation of millimeter wave propagation loss is spatially and temporally small because of the fewer multipaths, and therefore, the propagation loss has strong correlation with the user location. By using the decreased set of candidate beams, the beam search time can be reduced in the second stage. Then the beam refinement can increase the beamforming gain to increase user throughput in the third stage. To confirm the effects of the proposed beam search method, we conduct system level simulations by using a propagation model for millimeter wave communications proposed by MiWEBA, which is an international project between Europe and Japan. The results show that the proposed beam search method can reduce the number of candidate beams, and can therefore shorten the beam search time by about 39% without any degradation in outage probability compared with a conventional method.