Background Early language skills are critical for later academic success. Lower socioeconomic status (SES) children tend to start school with limited language skills compared to advantaged peers. We test the hypothesis that this is due in part to differences in caregiver contingent talk during infancy (how often the caregiver talks about what is in the focus of the infant's attention). Methods In a randomised controlled trial with high and low SES families, 142 11‐month olds and their caregivers were randomly allocated to either a contingent talk intervention or a dental health control. Families in the language intervention watched a video about contingent talk and were asked to practise it for 15 min a day for a month. Caregiver communication was assessed at baseline and after 1 month. Infant communication was assessed at baseline, 12, 15, 18 and 24 months. Results At baseline, social gradients were observed in caregiver contingent talk to their 11‐month olds (but not in infant communication). At posttest, when infants were 12 months old, caregivers across the SES spectrum who had been allocated to the language intervention group engaged in significantly more contingent talk. Lower SES caregivers in this intervention group also reported that their children produced significantly more words at 15 and 18 months. Effects of the intervention did not persist at 24 months. Instead expressive vocabulary at this age was best predicted by baseline infant communication, baseline contingent talk and SES. Conclusions A social gradient in children's communication emerges during the second year of life. A low‐intensity intervention demonstrated that it is possible to increase caregiver contingent talk and that this is effective in promoting vocabulary growth for lower SES infants in the short term. However, these effects are not long‐lasting, suggesting that follow‐up interventions may be necessary to yield benefits lasting to school entry.