Although gesturing onscreen instructors are widely included in video lectures, it is still unclear whether, when, and how they are conducive to learning. To clarify this issue, we conducted a set of three-level meta-analyses of 662 effect sizes from 83 articles, spanning Web of Science, PsycINFO, ERIC, Education Research Complete, ProQuest Dissertations & Theses, and Google Scholar up to March 2024. We included randomized controlled trials of gesturing instructors in multimedia learning, measuring retention test score, transfer test score, fixation time, fixation count, cognitive load, and/or social perception across all languages of publication. Funnel plot and Egger sandwich test were used to assess risk of bias. Results showed that adding gesturing instructors improved retention (g = 0.28, 95% CI:[0.19,0.37]) and transfer test scores (g = 0.31, 95% CI:[0.21,0.41]), yielding an "embodiment effect." This effect was stronger when the instructor displayed deictic, metaphorical, or a mixture of multiple gestures; when the instructor in the control condition was not visible; when the lecture was learner-paced and longer. Moreover, it increased learners' social connection ratings and eye fixation time and count on core learning material (but only when deictic gestures were used). Thus, gesturing onscreen instructors may promote learning by social and cognitive paths, deepening our understanding of the role of gesturing onscreen instructors in multimedia learning and providing guidance for designing effective video lectures. More studies with clear experimental descriptions and eye-tracking studies are needed.