Complementarities between partners’ characteristics are often held responsible for the patterns of assortative mating observed in marriage markets along different dimensions, such as race and education. However, when the marriage market is segmented into racially and educationally homogeneous clusters, people naturally have more match opportunities with their likes. In this paper, we build an empirically tractable dynamic matching model with endogenous separation and remarriage. In every period, agents participate in a competitive assignment game in the vein of Choo and Siow (2006), where mating strategies depend on both the expected match gains and search frictions in the form of meeting costs. We leverage panel data on the duration of both non-cohabiting and cohabiting relationships to jointly estimate both determinants of assortative mating with a nationally representative sample of the U.S. population. We show that, in the absence of search frictions, the share of matches between people of the same race (education) would decrease from 88.2% (49.2%) to 55.5% (40.8%), as opposed to 53.3% (33.5%) if singles were randomly matched. As a result, search frictions explain nearly all the racial homogamy observed in the data, but only approximately half of the observed educational homogamy, with the other half attributed to match complementarities. In a counterfactual exercise, we show that minority groups experiencing an unfavorable gender ratio when marriage markets are segmented, such as Hispanic men and black women, would benefit from access to a broader and more diverse pool of partners.