On the exact analytic visibility side:
Rather than casting a gazillion ray in the hope that N=gazillion is actually enough, there are researchers who have developed analytic algorithms, which will compute exactly which triangles, objects or cells are visible from another cell (i.e., the same result as if you case an infinite set of rays).
See Shaun Nirenstein's exact visibility page for a set of publications on this subject. You may note that one of the authors of the 2005 paper is the CTO of Umbra Sotware (the visibility middleware paired with Unity 3). I don't know whether or not this technology is used in the Umbra product or not. For scenes with specific structure (i.e., cell/portal), an analytic solution is covered nicely in Seth Teller's 1992 thesis here. Jiri Bittner's thesis also presents an analytic solution.
Generally, analytic visibility is hard to get right (mathematically complex, numerical issues, speed, etc.), but if done right it would make for a compelling piece of middleware.
On the ray casting side:
The algorithm you described has been around since the late 80's
(see Towards Image Realism with Interactive Update Rates
in Complex Virtual Building Environments).
There have been many advances in the ray-casting approaches in the last few years. Researchers have devised algorithims that choose fewer, but better rays using heuristics, and also may compute visibility over multiple cells simultaneously. See the papers on Peter Wonka's page (Guided Visibility Sampling and Adaptive Global Visibility Sampling). These get great results.
Probably the class of technique that is still used the most today in practice is the GPU accelerated hemicube algorithms. These just sample visibility from a set of points by rendering a 6 sided cube map and reading back object/triangle/cell indices as data. Occlusion queries can also be used to reduce the amount of data read back. Algorithms based off this tend to be easy to implement and give good results in practice (fast as hell, since they use the GPU), but the sampling distribution (millions of rays cast from a realtively small set of single points), may make it take longer to converge on an accurate solution then one would think. This technique tries to choose better visibility cube source locations than a simple uniform or random sampling.
There are plenty of other visibility techniques (hundreds). I suggest reading some of the background chapters in the theses linked to, if you want to know more.