Although eligibility for antiretroviral treatment is no longer based on CD4 thresholds, CD4 testing remains important. Variation in CD4 cell count complicates initiation of antibiotic prophylaxis, differential diagnoses and assessments of immunological recovery. Five hundred and fifty-three HIV-positive antiretroviral-naïve adults, recruited from inner-city clinics, had three serial CD4 cell count tests. Test 1 was mostly done in a laboratory network supporting primary care clinics, while Tests 2 and 3 were performed in a tertiary-level laboratory. Reproducibility was assessed through Bland–Altman limits of agreement and coefficients of variation. Participants, a mean age of 34 years and mostly female (57%), had a median 203 CD4 cells/μL (Test 1). Seventeen per cent classified as having advanced HIV disease (CD4 cell count < 200 cells/µL) on Test 1 had a CD4 cell count > 200 cells/µL on Tests 2 and 3. Mean differences between tests were <10 cells/µL for all comparisons. Limits of agreement for Tests 1 and 2 were −106.9 to 112.7 and coefficient of variation 15. Corresponding figures for Tests 2 and 3 were −88.2 to 103.4, and 13. Means of tests were similar, suggesting no systematic measurement differences, despite testing being done at different times. Variations were, however, considerable in many instances, though smaller in testing done in the same laboratory. CD4 cut-offs must not be applied rigidly, but rather constitute one amongst many factors used to guide patient care. Moreover, given the difficulties in determining whether CD4 changes are due to HIV disease, or other biological and laboratory factors, CD4 laboratory reports should include a range of plausible values, not only the absolute count.