Mild cognitive impairment (MCI) is a neurocognitive disorder which involves cognitive impairments beyond those expected based on an individual’s age and education, but which are not significant enough to interfere with instrumental activities of daily living. Reading, combined with eye-tracking, has been proposed as a sensitive biomarker to detect MCI and differentiate it from healthy controls. The primary objective of this study was to investigate the differences in eye movements during reading between people with MCI and healthy controls. Additionally, on the basis of observed differences, we designed features to be used in a model for automatic detection of cognitive impairment. A hundred and fifteen (115) subjects were enrolled in the study. All subjects underwent a neurological examination, an extensive psychological assessment, and a short reading exercise during which their gaze behaviour was recorded with an eye-tracker. Based on the neuropsychological assessment, the subjects were divided into two groups: cognitively impaired (n = 62) and healthy controls (n = 53). Their gaze behaviour during reading was described in terms of features based on saccades and fixations and compared between the two groups. A number of features exhibited very high statistical significance showing that gaze behaviour is quite different between the observed groups. These features were used to construct a model using machine learning to detect cognitively impaired subjects from healthy controls. The model exhibits promising results of nearly 80% area under the receiver operating characteristic curve, however complementary eye-tracking exercises could further increase its detection accuracy.