Although labor market "mismatch" often refers to an imbalances in supply and demand across occupations, mismatch within occupations can arise if skill requirements are changing over time, potentially reducing aggregate matching efficiency within the labor market. To test this, we examine changes in employer education and skill requirements using a database of 200 million U.S. online job postings between 2007 and 2019. We find that the degree of persistence in educational upskilling lasted longer than was previously known and was not uniform but rather varied considerably across occupations and was often coupled with an increased demand for software skills. We also find evidence that upskilling contributed to reduced matching efficiency in certain segments of the US labor market as well as in the aggregate. In particular, matching efficiency was lower in higher-skilled occupations, potentially because they are becoming more specialized, and possibly explaining growing wage polarization and inequality.