Graph neural networks (GNNs) deliver state-of-the-art results in node classification, link prediction, and whole-graph prediction across chemistry, recommender systems, and social-network analysis. However, training them demands large labelled graphs and extensive GPU time, so a stolen model represents a costly loss of intellectual-property (IP) investment while retraining is prohibitively expensive. To counter this risk, we present EGNNFingers—a task-agnostic, explainability-driven fingerprinting framework expressly designed for GNN IP protection. EGNNFingers employs attribution methods to distil the relationship between inputs and outputs into a compact fingerprint that uniquely characterises the source model. Owners can later issue a small set of graph queries to an online service and, from the returned explanations, decide whether the deployed model originated from their protected GNN. Experiments on three public benchmarks show that EGNNFingers consistently achieves higher ownership-verification accuracy than the original GNNFingers method, confirming its practical value for protecting GNN IP and detecting piracy in real-world deployments.