This study examined the mechanism of effect of the WHO Caregiver Skills Training (CST) through secondary analysis of a pilot RCT conducted in community settings. Participants were 86 caregivers (77% mothers) of children with ASD (78% male, mean age: 44.8 months) randomized to CST (n = 43) or treatment as usual (n = 43). The primary outcomes, measured at baseline (t1), immediately post‐intervention (t2), and 3 months post‐intervention (t3), were derived from the coding of caregiver‐child free play interactions with the Brief Observation of Social‐Communication Change (BOSCC) and the Joint Engagement Rating Inventory scale (JERI). At t3 positive treatment main effects had been observed for caregiver skills supportive of the interaction and for flow of the interaction (JERI), albeit only non‐significant changes in the expected direction for child outcomes: autism phenotypic behaviors (BOSCC), joint engagement and availability to interact (JERI). This study tested the theory of change of CST, hypothesizing that the intervention would lead to an improvement on all child and dyad outcomes through an increase in the caregiver skills supportive of the interaction. Serial mediation analyses revealed that the effect of the intervention was significantly influenced by change in caregiver skills. Participation in the intervention led to notable increases in caregiver skills at t2 and t3, which subsequently contributed to improvements at t3 in flow of the interaction, autism phenotypic behavior, joint engagement, and availability to interact. We confirmed our a priori hypothesis showing that change in caregiver skills significantly mediated the effect of treatment on the dyad primary outcome, as well as on the other child outcomes that had shown non‐significant changes in the expected direction. Implications for intervention design and policy making in the context of public health services are discussed.