A gulf exists between prior work testing metacognitive instructional interventions and teacher practices that may support metacognition in the classroom. To help bridge this gulf, we designed an observational protocol to capture whether and how teachers provide metacognitive support in their talk and examined whether these supports were related to student learning. We examined four features of metacognitive support, including the type of metacognitive knowledge supported (personal, strategy, or conditional), the type of metacognitive skill supported (planning, monitoring, or evaluating) the type of instructional manner in which the support was delivered (directives, prompting, or modeling), and the type of framing (problem specific, problem general, or domain general), during three types of instructional activities (individual, group, or whole-class instruction). We compared teacher talk from 20 middle school mathematics classrooms with high growth in conceptual mathematics scores with teacher talk from 20 classrooms with low growth. For each of these classrooms, we examined the amount of teacher talk that supported metacognition during one regular class period. Observations revealed that the high-conceptual growth classrooms had more metacognitive supports for personal knowledge, monitoring, evaluating, directive manners, and domain-general frames than the low-conceptual growth classrooms. We discuss the implications of those observations for bridging research on metacognition to teacher practice.