You can do it either way, but the settup is very different for each.
If you want to AC couple a digital 1 or 0 onto the DC VCO tune voltage, then you need the loop bandwidth to be very much smaller than the data rate. This way the PLL keeps the average frequency of the VCO on the center frequency desired, but allows the instantaneous frequency to swing +/- with the data. Two caveats:
1) you need a big divisor ratio (VCO divided by N) to keep the phase detector swing <<0.05 radians, and 2) you might need a randomizer on the digital data, since if you send a string of all 0's, the frequency is going to eventually drift back to the middle, and then when you finally send a 1, it is going to jump twice too far the other way.
The second way is to have a very high control loop bandwidth, much bigger than the data rate. Then you modulate the signal by changing the divisor ratio on the PLL. For instance, with a 10 KHz reference frequncy, if your VCO frequency desired is 100 MHz, you would divide by 10,000. If you change the divisor between 9,999 and 10,001, you would have frequency modulation of +/- 10 KHz.