I draw the line by budgeting phase noise by offset, not by buying the lowest phase-noise part I can afford. As LunaWei noted, inside the PLL loop bandwidth the reference dominates, and outside it the VCO dominates — so a very clean 10 MHz reference does not help if your VCO is the bottleneck at 100 kHz offset. I start by writing down the carrier, channel bandwidth, EVM and ACLR limits, then the multiplication factor, because reference phase noise rises by 20 log N when multiplied.
For a typical sub-6 GHz small cell I have usually been fine with a good TCXO, not an OCXO, provided the loop bandwidth is chosen well and the VCO is decent. The OCXO only earns its cost when temperature stability or holdover matters — for example, an outdoor unit that must stay locked through a GPS outage, or a design where aging drift would push the loop out of its clean region. If GPS or PTP disciplining is available, a TCXO plus disciplining is often the cheaper, safer trade.
The acceptance test that has worked for me is a before-and-after swap: measure EVM and ACLR with the candidate reference, then with a better reference, at the same output power and temperature. In one small-cell build the OCXO improved integrated jitter by about 30%, but EVM moved by less than 0.2%, so the extra cost was not justified. In another, the reference swap fixed a close-in spur that was failing ACLR — there the OCXO paid for itself. Which band and loop bandwidth are you using, and is holdover a hard requirement for your deployment?