Statistical mediation analysis is both contested and widely used in psychology for testing theories about causal processes. First, I present a scoping review of mediation analyses from 3,889 eligible articles published between 2023 and 2024 in leading psychology journals. Relative to past reviews, some improvements were evident—larger samples (median N = 402) and more frequent testing of multiple mediators or alternative models (56% vs. 9-10% in 2015). However, most reported mediations lacked tests of moderation (71%) or statistical power (86%), with a third (38%) using solely correlational data (18% fully cross-sectional). Reporting was inconsistent: Some papers omitted component paths (17%), total effects (11%), or direct effects (22%), even though reliance on index-only approaches can mask model misspecification. Next, I extend models of causal heterogeneity in mediation by identifying a new concern: illusory indirect effects arising from omitted moderators. Using simulations I demonstrate these illusory indirect effects (which satisfy popular mediation criteria), then provide user-friendly diagnostics to detect this problem in existing datasets. A new individual-level index of mediation—derived from standard regression diagnostics—quantifies causal heterogeneity within samples, offering a data-driven method for identifying overlooked moderators. I conclude by discussing practical implications for advancing psychological theory across diverse populations.