In the test-pair similarity effect, forced-choice recognition is more accurate for similar test pairs, such as leopard-cheetah, than it is for unrelated test pairs, such as leopard-turnip. According to global matching models, this occurs because the retrieved familiarities of similar items are correlated. In the Minerva 2 model, global matching underlies frequency judgments as well as recognition memory. One implication of this model is that judged frequencies of similar items should be correlated. Another implication is that judgments of summed frequency for pairs of words (how many presentations were there of word1 and word2 combined?) should have higher variance when word1 and word2 are similar than when they are unrelated. These predictions were tested and confirmed in two experiments. A review of these and other results suggests that theories of recognition memory should also be applicable to frequency-judgment tasks.