"Subtle" is a good word for it. The biggest problem with this methodology is that the signal is never clean -- there's random electrical noise both in all the written data (the write heads have noise in their signal when originally laying down data) and in the read pick-up heads.
There are two basic methodologies:
1. Magnetic Remnance
This works on the principle that different patterns of 0 and 1 will result in slightly different magnetic values. The problem is that the more a drive has been used and the more it is overwritten, the more subtle these differences will become, and that harder to detect.
The situation is further complicated by not knowing the overwriting sequence or which "layer" you are interested in.
It's also been noted that actual investigations find most of the track-edge signal is actually switching noise from the write heads (from the overwritten and overwriting data), not the data you're looking for.
2. Track Overlay
This is based on the principle that the heads don't write to exactly the same place on each "pass" so some residual data remains at the edges. Once again the greater the number of overwriting passes, the more difficult it is to recover anything useful.