My (simple) noise model of the AD5066 is an 8k resistor's thermal noise plus the reference buffer's thermal noise (~60nV/rt-hz ??) multiplied by code/(full-scale code).Does that sound reasonable?
Noise from the reference will couple to the output so the plot would include reference noise.
The AD55541A or AD5542A have better noise numbers but you will need to add an amplifier on the output.