PDP shows the marginal average effect of feature(s) on the target variable so it helps us understand how different features impact (e.g. directionality of the correlation; different slopes of PDP plot in different value intervals tell us the degree of impact) the target variable. But as mentioned in the article, the assumption that the feature of interest is not correlated with other features should be met which is quite tricky in real-world datasets. There are tons of other XAI methods that have different pros and cons so which method to choose really depends on what goal you have and what circumstance you are in. If you more interested in "causal" inference, more econometrics oriented methods that try to control for observed and unobserved confounding features such as difference-in-difference, instrumental variables, and regression discontinuity design are available.
I would highly recommend this article that a senior researcher at Microsoft wrote which warns misleading interpretations of XAI methods to prove causality (he used SHAP as an example but it's pretty much applicable to any XAI methodology).
