It is difficult to test near-infrared spectrophotometry instruments in vivo. Therefore we constructed a liquid phantom which mimics the neonatal head. It consists of a spherical 3.5 mm thick layer of silicone rubber simulating skin and bone and a 0.5 mm thick clear layer of polypropylene imitating cerebrospinal fluid. It acts as container for a liquid solution with Intralipid, 60 micromol l(-1) haemoglobin and yeast. The solution was oxygenated using oxygen and then deoxygenated by the yeast. From the instrumental (Critikon 2020) algorithm, we found that with increasing scattering (0.5%, 1%, 1.5% and 2% Intralipid concentration) the reading was increasingly offset from the expected value of 0 micromol l(-1) by 55.7, 68.6, 76.5 and 80.4 micromol l(-1) (oxyhaemoglobin) and 16.0, 24.4, 29.6 and 31.7 micromol l(-1) (deoxyhaemoglobin). This reduced the range of the oxygen saturation reading from the expected 100% to 31.5, 21.1, 14.3 and 11.5%. Haemoglobin concentration changes were increasingly underestimated by a factor of two to four. For a second algorithm based on the diffusion approximation the offsets were smaller: oxyhaemoglobin 11.4, 17.8, 22.5 and 25.1 micromol l(-1) and deoxyhaemoglobin 1.3, 3.4, 5.2 and 6.0 micromol l(-1). The range of the oxygen saturation reading was higher: 41.3, 29.2, 23.4 and 16.6%. Concentration changes were underestimated by a factor of six to ten. This study demonstrates the need to develop algorithms which take into consideration anatomical structures.