Choosing the number of clusters combines several criteria: the inertia elbow, the silhouette score, the stability of clusters between runs, and above all business usefulness. There is no universally correct k; the choice must be justified and the clusters must remain interpretable.