Research on educational Augmented Reality (AR) began approximately three decades ago, starting in the early 1990s. Since then, numerous systematic reviews and meta-analyses have synthesized evidence regarding its impact on learning. However, these analyses have reported mixed results, ranging from a small to a medium impact. Mixed results are common, especially in meta-analyses, as this type of study cannot eliminate sampling error variance. Consequently, their conclusions may be affected by sample bias and publication bias. This study presents a second-order meta-analysis to identify more reliable insights into the impact of AR in educational settings. For this study, we synthesized 19 eligible articles, incorporating 21 effect sizes from first-order meta-analyses that explored the impact of AR in comparison to non-AR control groups. These meta-analyses encompassed 657 effect sizes derived from 273 primary studies published between 2000 and 2022. Furthermore, we analyzed the moderating effects of domain discipline, participants, and the type of learning outcome. Quantitative analysis revealed that AR has a medium impact on learning outcomes (g=0.64,p<.001; 95% CI [0.56,0.72]). Contrary to our expectations, the findings of this study indicate that none of the three factors - discipline, participants, and type of learning outcome - moderated the effect of AR on learning. This implies that the impact of AR is consistent across disciplines, participant characteristics, and outcome measures. The study concludes that AR has been a versatile pedagogical alternative in education and can successfully complement educational processes. Finally, the study explains the limitations of the results and highlights new directions for future research.