The spread of fake news became a social issue following the 2016 U.S. presidential election, making the detection of fake news an urgent task. To address this, detection methods using machine learning models based on linguistic features have been proposed. However, these methods have been evaluated only on specific datasets, and their reliance on linguistic features poses challenges for generalizability. In this study, we propose a fake news detection model that does not rely on linguistic features. We use a large language model to select ten types of features—such as claim substantiation, source credibility, and sensational expressions—and to evaluate the degree of fake news associated with each feature. We compare the proposed model with a model that relies only on linguistic features. As a result, we demonstrate improved generalization performance on unseen datasets. On the other hand, an analysis of the features that contributed to fake news detection revealed variability across datasets, indicating that further investigation is needed to improve detection accuracy.