The therapeutic alliance has been found to predict psychotherapy outcome in numerous studies. However, critics maintain that the therapeutic alliance is a by-product of prior symptomatic improvements. Moreover, almost all alliance research to date has used differences between patients in alliance as predictor of outcome, and results of such analyses do not necessarily mean that improving the alliance with a given patient will improve outcome (i.e., a within-patient effect). In a sample of 646 patients (76% women, 24% men) in primary care psychotherapy, the effect of working alliance on next session symptom level was analyzed using multilevel models. The Clinical Outcomes in Routine Evaluation–Outcome Measure was used to measure symptom level, and the patient version of the Working Alliance Inventory–Short form revised (Hatcher & Gillaspy, 2006) was used to measure alliance. There was evidence for a reciprocal causal model, in which the alliance predicted subsequent change in symptoms while prior symptom change also affected the alliance. The alliance effect varied considerably between patients. This variation was partially explained by patients with personality problems showing stronger alliance effect. These results indicate that the alliance is not just a by-product of prior symptomatic improvements, even though improvement in symptoms is likely to enhance the alliance. Results also point to the importance of therapists paying attention to ruptures and repair of the therapy alliance. Generalization of results may be limited to relatively brief primary care psychotherapy.