I think it would actually be more evenly distributed with a global counter or a session table entry.
As there will be more TCP connections with a single HTTP request, resetting the counter on the start of every connection would result in many more requests going to the first redirect location. If you didn't want to use a global counter or session table entry, you could use an idea from Spark where you create a pool with two dummy pool members, force a load balancing decision with LB::select and then choose the redirect location based on which dummy pool member was selected.
If any one of these options sounds good, let us know and we can provide an example.
Aaron