It used to be the gate length. But eventually improvements were made that didn't neatly map to the gate length but still doubled transistor density, so they transitioned to just dividing the number by sqrt(2) each generation (since chips are two-dimensional making everything sqrt(2) smaller in each dimension increases density by factor 2).
Until 14nm the relation to transistor density was kind of true, all 14nm processes have similar density that is roughly double that of 22nm, which is roughly double that of 32nm. But now even that's meaningless and all they're doing is steps of factor sqrt(2) out of habit.
Until 14nm the relation to transistor density was kind of true, all 14nm processes have similar density that is roughly double that of 22nm, which is roughly double that of 32nm. But now even that's meaningless and all they're doing is steps of factor sqrt(2) out of habit.