Instead of computing the distance to the centroid of the area
compute the distance of the area to the centroid of the feature.
This means we give preference to the area that covers the centroid.
It's still a heuristics but one that is a bit less random.