I've considered various ways, e.g. recursive subdivision with raytesting, rendering a coarse version of occluders to see what they cover. Feel free to suggest any other approaches.

Do you need exact results, or just good enough to optimise your rendering?

Once approach that I've seen in a few games that I thought was so simple that it's clever is:
*Project the conservative bounds of all your occluders into screen-space.
*Select the 5 (or n) occluders with the largest screen area.
*Create a frustum from each of these occluders to get their occlusion volume.
*Perform a "frustum fully contains" test for each occludee's bounding volume against each occluder frustum.

It's very dumb, coarse, and kind-of brute-force, but it's been used in a few AAA console games and is actually very efficient computation-wise (assuming you've got sensible data layouts).

I've considered various ways, e.g. recursive subdivision with raytesting, rendering a coarse version of occluders to see what they cover. Feel free to suggest any other approaches.

Do you need exact results, or just enough to optimise your rendering?
Once approach that I've seen in a few games that I thought was so simple that it's clever is:
*Project the bounds of all your occluders into screen-space.
*Select the 5 (or n) occluders with the largest screen area.
*Create a frustum from each of these occluders to get their occlusion volume.
*Perform a "frustum fully contains" test for each occludee's bounding volume against each occluder frustum.

It's very dumb, coarse, and kind-of brute-force, but it's been used in a few AAA console games and is actually very efficient computation-wise (assuming you've got sensible data layouts).