Recommender systems (RSs) are AI tools used in online platforms to algorithmically identify and promote to each platform’s user a small number of personalised items, such as, news, posts, music, and videos. Their core component is a machine learning algorithm exploiting sparse users’ online behaviour data to estimate what items, in a predefined catalogue, should be presented to each user to achieve a desired goal (e.g., increase revenues and user satisfaction). After having briefly discussed RSs value, and risks, we will focus on the role and importance of a proper system evaluation, which is also a mandatory request formulated in the European legislation (Digital Service Act). Classical machine learning evaluation approaches, which are based on offline testing the predictive model on holdout data, are insufficient to asses the true effect of a new RS algorithm (intervention). A new type of simulation-based evaluation approaches will be therefore introduced: they enable to estimate the potential effects (positive and negative) of an RS on the choices made by their users when they are influenced by the received recommendations. The application of this evaluation method will be exemplified in a particular case: sustainable and multistakholder recommendations in tourism, namely, how to tame over tourism and respect local communities. We have found that, under certain conditions, multiple stakeholder can jointly benefit from the RS, hence creating a more stable and trustful cooperation scenario.