In mathematical word problem solving, reading and mathematics interact. Previous research used the method of eye tracking to analyze reading processes but focused on specific elements in prototype word problems. This makes it difficult to compare the role of reading in longer, more complex word problems and between individuals. We used global measures of eye movements that refer to the word problem as a whole, similar to methods used in research on eye movements during reading. Global measures allow comparisons of reading processes of word problems of different structure. To test if these global measures are related to cognitive processes during word problem solving, we analyzed the relation between eye movements and the perceived difficulty of a task and its solution rate. We conducted two experiments with adults and undergraduate students (N = 17 and N = 42), solving challenging mathematical word problems from PISA. Experiment 1 showed that more difficult items were read with shorter fixations, more saccades, more regressions, and slower, with correlations ranging from r = 0.70 to r = 0.86. Multilevel modelling in experiment 2 revealed that for the number of saccades and the proportion of regressions, the relationship was stronger for low-performing students, with performance explaining up to 37% of the variance between students. These two measures are primarily associated with building a problem model. We discuss how this approach enables the use of eye tracking in complex mathematical word problem solving and contributes to our understanding of the role of reading in mathematics.