Improving teacher evaluation is one of the most pressing but also contested areas of educational policy. Value-added measures have received much of the attention in new evaluation systems, but they can only be used to evaluate a fraction of teachers. Classroom observations are almost universally used to assess teachers, yet their statistical properties have received far less empirical scrutiny, in particular in consequential evaluation systems. In this essay, we highlight some conceptual and empirical challenges that are similar across these different measures of teacher quality. Based on a review of empirical research, we argue that we need much more research focused on observations as performance measures. We conclude by sketching out an agenda for future research in this area.