Actually the Z occlusion query can significantly help reduce the cost of fill rate as well as well as geometry processing.
The bounding box only requires a single z test per pixel across its screen projection. The number of z tests per pixel for the geometry inside the bounding box is equal to its average depth complexity.
You do not get any fill rate improvement when the bounded geomtery consists of convex objects with an average (front facing) depth complexity of 1, such as a sphere. This was presumably the case mentioned above.
However, if the bounded geometry is highly complex, such as a complete 3d tree with an average front facing depth complexity of say 8, then the fill rate improvement would be significant. Of course, the improvement in geometry processing would also be significant in this case.
The bounding box only requires a single z test per pixel across its screen projection. The number of z tests per pixel for the geometry inside the bounding box is equal to its average depth complexity.
You do not get any fill rate improvement when the bounded geomtery consists of convex objects with an average (front facing) depth complexity of 1, such as a sphere. This was presumably the case mentioned above.
However, if the bounded geometry is highly complex, such as a complete 3d tree with an average front facing depth complexity of say 8, then the fill rate improvement would be significant. Of course, the improvement in geometry processing would also be significant in this case.