When and why randomised exploration works (in linear bandits)