If you want to throw machine learning at it, the browser could try to predict what site the user thinks they are on, and highlight any discrepancies.

The data collection necessary to get an accurate model would however be quite invasive, so I'm not quite sure whether I'd use a browser with that ability. Maybe just having visual fingerprints of the most visited sites and serving them as a static model would be enough.