Causal Structure Learning (CSL), amounting to extracting causal relations among the variables in a dataset, is widely perceived as an important step towards robust and transparent models. Constraint-based CSL leverages conditional independence tests to perform causal discovery. We propose Shapley-PC, a novel method to improve constraint-based CSL algorithms by using Shapley values over the possible conditioning sets to decide which variables are responsible for the observed conditional (in)dependences. We prove soundness and asymptotic consistency and demonstrate that it can outperform state-of-the-art constraint-based, search-based and functional causal model-based methods, according to standard metrics in CSL.