Purpose: This study aims to help researchers design observational measurement systems that yield sufficiently stable scores for estimating caregiver talk among caregivers of infant siblings of autistic and non-autistic children. Stable estimates minimize error introduced by facets of the measurement system, such as variability between coders or measurement sessions. Method: Analyses of variance were used to partition error variance between coder and session and to derive g coefficients. Decision studies determined the number of sessions and coders over which scores must be averaged to achieve sufficiently stable g coefficients (0.80). Twelve infants at elevated likelihood of an autism diagnosis and 12 infants with population-level likelihood of autism diagnosis participated in two semistructured observation sessions when the children were 12-18 months of age and again 9 months later. Caregiver follow-in talk was coded from these sessions. Results: Two sessions and one coder were needed to achieve sufficient stability for follow-in talk and follow-in comments for both groups of infants at both time points. However, follow-in directives did not reach sufficient stability for any combination of sessions or coders for the population-level likelihood group at either time point, or for the elevated likelihood group at Time 2. Conclusion: Researchers should plan to collect at least two sessions to derive sufficiently stable estimates of caregiver talk in infants at elevated and general population-level likelihood for autism.