Objective: In our study, we investigate the effect of data quality on the accuracy of models that predict users’ subjective impressions from visual appearances of websites. Background: Machine Learning (ML) methods are increasingly employed to predict user behavior, particularly subjective impressions that cannot be automatically inferred from user interfaces (UIs) or from interaction logs. In this field, feature-based ML methods are often utilized instead of data-hungry deep learning ones, since collecting training data from human subjects requires considerable effort. The features for the models can be automatically extracted from UI code or screenshots with dedicated algorithms, which are an active area of research. Still, relying on human UI labelers is often considered a better option, since they presumably provide more accurate input data that allows more accurate prediction. Method: For about 500 homepages of universities’ and colleges’ websites, quantitative values for several UI features were obtained: number of UI elements, share of whitespace, etc. The values came from 11 human labelers of varying diligence, whose work quality was validated by another 20 verifiers. The dependent variables in the models were subjective impressions from the web pages, operationalized as the Likert scales of Complexity, Aesthetics and Orderliness, assessed by 70 subjects. Results: We built 33 linear regression models and analyzed their capability (as represented by R²s) to predict the subjective impressions. Conclusion: Although many research works propose new features for UIs and/or improvement of algorithms that calculate them, the discovered effect of the input data quality on the prediction accuracy was scale-dependent and, somehow unexpectedly, negative. Application: The results of our study might aid in balancing data quality and data quantity in projects that use ML methods to predict subjective impressions of website users.