As far as I can tell, there is no specific rule. It will depend in part on how crowded your scene will become with items that you want to detect and locate separately. Creating a high granularity grid increases computational cost for training, and there is no reason to do so if it would only cover additional cases that are much rarer than the detection accuracy that the base algorithm achieves.
The choice can be driven by the ground truth data. The ground truth for YOLO needs to be expressed in the form of grid locations with classes and bounding rectangle sizes. If you don't find any training examples where you want to label two items with their centre inside the same grid square, then this is a good indication that your grid size is fine-grained enough.
Even if there are one or two examples with a clash like this, you may be able justify labelling just one item in the ground truth and be OK with a resulting model that is not able to cope with close overlap between two separate objects. And even with smaller grid squares, YOLO may not be able to learn to separate the objects, if such an overlap only occurs rarely.
I would expect a simple rule of diminishing returns applies. As datasets grow larger, and object detection can be trained on more powerful computers, we may see state of the art models still using YOLO but with more grid points.