The process of quantifying trust considers the factors that affect it, which can be applied to identify malicious behavior, reduce uncertainty, and facilitate decision-making. Traditional trust evaluation methods based on statistics and reasoning, rely heavily on domain knowledge, which limits their practical applications. Graph Neural Networks (GNNs) are a new Machine Learning (ML) paradigm that can revolutionize the evaluation of trust, by modeling relationships as graphs to simplify relevant data and automating end-to-end evaluation. Thus, a variety of GNN-based trust evaluation models have been developed for different applications. However, there is still a gap in the literature regarding a review on these advances with discussion about remaining challenges. To bridge this gap, we conduct the first review on GNN-based trust evaluation. We first propose a set of criteria in terms of trust-related attributes, correctness, functionality, and overhead. Then, we propose a taxonomy of existing GNN-based trust evaluation models, followed by a review using the proposed criteria to analyze their pros and cons. A quantitative analysis of the recent cutting-edge models is also provided. Based on the review and experimental results, we identify key challenges and suggest future research directions.