No ground truth to check against
With supervised learning, evaluation is relatively straightforward: you hold out some labeled examples, run the model on them, and compare its predictions to the known correct answers. Accuracy, precision, recall, these all rely on having a ground truth to measure against. Unsupervised learning has no such luxury. When k-means produces five clusters, there's no labeled answer key saying whether five is right, or whether the specific points it grouped together truly belong together in any objective sense.
This isn't a minor inconvenience, it's a structural feature of the problem. Unsupervised learning was chosen specifically because labels don't exist. So by definition, there's no way to score the output against a known correct grouping the way you'd score a spam classifier against emails a human already labeled.