The instructor's eye gaze can serve as an important social cue in video lectures. The current study used two sets of three-level meta-analyses to explore the effects of the instructor's guided gaze or the instructor's direct gaze on learning outcomes, fixation time, perception of parasocial interaction, and cognitive load. A total of eight meta-analyses [2(eye gazes)×4(dependent variables)] were included. Eighteen studies with a total of 203 effect sizes were identified. The results showed that guided gaze significantly promoted learning outcomes [g = 0.33; guided gaze vs. no guided gaze (i.e., direct gaze, averted gaze, or no gaze)], and direct gaze significantly promoted learning outcomes [g = 0.30; direct gaze vs. no direct gaze (i.e., averted gaze or no gaze)], significantly increased perception of parasocial interaction (g = 0.34), and significantly reduced fixation time on the learning material (g = -0.65). Moderating effect analyses showed that learning outcomes of fixed guided gaze (g = 0.57; instructors look at the instructional screen) were significantly better than that of shifting guided gaze (g = 0.27; instructors switch their eye gaze between the instructional screen and camera). Learning outcome effects with a control group with averted gaze (g = 0.76) were significantly higher than those with direct gaze (g = 0.32) or no gaze (g = 0.22). This study suggested that guided gaze and direct gaze have different effects on learning. In practical teaching, instructors should use guided gaze and direct gaze, while avoiding averted gaze and no gaze.