We studied the reproducibility of metabolite signals (from N-acetyl aspartate [NAA], choline, and creatine) measured with a standard single-voxel proton magnetic resonance spectroscopy technique (PRESS, TE = 135 ms, 8 ml VOI) in vitro and in two groups of normal volunteers. Spectral peak areas were quantified both by integration and by curve-fitting. In the in vitro study, the "between-days" variability (coefficient of variation [CV]) of measurements ranged from 0.9% to 2.3%. In the first group of volunteers (n = 12), single voxel spectroscopic measurements (8 ml VOI, 256 acquisitions [ACQs]) were made from mirror-image parts of the right and left hemispheres on 2 separate days. The "between-days" CV of measurements ranged from 9% to 18% for metabolite areas, and from 10% to 26% for metabolite area ratios. There were no significant differences between quantification method or hemisphere. After checking and optimising the MR scanner performance (in fact, it was virtually optimal), the second group (n = 4) each had six sequential single voxel spectroscopic measurements (each of 64 ACQs) from the right hemisphere (without moving the voxel) on each of 4 separate days. Even when the metabolites were measured from the same place in the same hemisphere sequentially six times in a 20-min period, the "within-run" CVs ranged from 4.4% to 17.2% for metabolite areas and from 9.7% to 17.0% for metabolite area ratios. The between-days CVs for the subjects ranged from 7.7% to 25.8% (metabolite areas) and from 10.1% to 22.6% (metabolite area ratios). The variability is due to a combination of random noise, subject motion, baseline artefacts in the spectra, and uncertainties in repositioning the VOIs. It is likely to represent the best reproducibility possible with 8-ml VOIs in cooperative, healthy volunteers carefully positioned on each occasion in a standard clinical scanner. Changes in metabolite levels in individuals must therefore be of the order of 20-40% before we can be reasonably confident of measuring them. Reproducibility in patients, who may be less cooperative, will probably be no better, and this must be taken into account in the interpretation of MRS studies in patients with brain pathology; for example, stroke, head injury, and tumours.