We propose a test procedure for distinguishing between two separable classes of arbitrary distribution tails and, moreover, prove its consistency. The method is based on the goodness-of-fit test for an arbitrary distribution tail, and properties of this test are also discussed. This is the first goodness-of-fit test for distribution tails proposed in the literature that can be applied for testing the goodness-of-fit hypothesis with a discrete distribution tail. In contrast to the overwhelming majority of works related to statistics of extremes, we do not assume that the distributions belong to any of maximum domains of attraction.