An untargeted method using headspace solid-phase microextraction coupled to electronic nose based on mass spectrometry (HS-SPME/MS-eNose) in combination with chemometrics was developed for the discrimination of oranges of three geographical origins (Italy, South Africa and Spain). Three multivariate statistical models, i.e. PCA/LDA, SELECT/LDA and PLS-DA, were built and relevant performances were compared. Among the tested models, SELECT/LDA provided the highest prediction abilities in cross-validation and external validation with mean values of 97.8% and 95.7%, respectively. Moreover, HS-SPME/GC-MS analysis was used to identify potential markers to distinguish the geographical origin of oranges. Although 28 out of 65 identified VOCs showed a different content in samples belonging to different classes, a pattern of analytes able to discriminate simultaneously samples of three origins was not found. These results indicate that the proposed MS-eNose method in combination with multivariate statistical analysis provided an effective and rapid tool for authentication of the orange's geographical origin.