they're not segregated per se, but they are effectively contiguous media regions. like a phone book, the 'a's are segregated from the 'b's but only because of perception - its really just one big long book. This assumes that you consider a cylinder to be a logically contiguous area, even though by definition a cylinder is not physically contiguous space because it spans platters.
Altho, there is no reason that a drive must be implemented this way - after all, technically the drive could use a spiral notebook with bits written in crayon as long as it looked right from the outside.
Evidence of this being the actual implementation used can be seen by noting the large performance difference between partitions stored toward the outside edge of the platters and those on the inner cylinders. If you were to find a drive that did not exhibit this behavior, then that would imply some other implementation. If for some reason consistent performance across the drive were paramount, then an alternative implementation (such as one that involved distributing the data in blocks across the entire drive, or perhaps even just aggregating by tracks instead of by cylinders) could be useful. I have never seen such a drive, most likely because aggregation of cylinders means that the heads don't have to be repositioned as often (imagine a partition that occupied a single cylinder's amount of space: with no head movement at all, the entire partition is accessible); the tradeoff (for a drive with a constant-speed spindle) is that some of your partitions are relatively fast and some are relatively slow.