Abstract Digital mental health interventions differ in how therapeutic content is organized or generated and delivered. We searched five databases through 31 May 2026 and included 404 randomized studies involving 110,037 participants. Twelve Bayesian random-effects network meta-analyses compared intervention architecture, delivery modality and their combinations for depression and anxiety at post-intervention and follow-up. The pooled direct effect across outcomes and stages favored digital interventions over primary controls (Hedges’ g = 0.338, 95% credible interval 0.306–0.371). Web ranked highest at post-intervention and App at follow-up, although web–app contrasts were uncertain. Large language model-based interventions ranked favorably, with limited follow-up evidence. Comparative rankings should be interpreted cautiously because most networks showed evidence of inconsistency and confidence in comparisons with Standard care was limited. Digital interventions showed modest symptom benefits, but the comparative evidence does not establish a preferred architecture or delivery modality.