Learning can be made more efficient when learners generate the to-be-learned text contents instead of passively receiving them. A multi-level meta-analysis was conducted to provide an overall estimate of the text generation effect's magnitude and to identify theoretically and practically relevant moderators. Overall, generation interventions improved learning with texts compared to reading them (Hedges' "g" = 0.41). This benefit was not attributable to time-on-task and was found across several learning conditions and settings (e.g., narratives and expository texts, multiple generation, and learning assessment tasks). The meta-analysis further suggests that generation benefits learning most strongly if the cognitive processes stimulated by the generation task complement those processes already stimulated by the text. In sum, the findings suggest that text generation can be suitable for educational applications especially if certain conditions are observed.