Also, if the part is an amplifier then you will have to look at the power supply regulation under two tone conditions to make sure you don't generate AM terms that will look like IMD.
You should be able to rule out the analyser as long as you use a good quality attenuator ahead of it and set the analyser's own attenuator carefully.
Also, can you find a signal combiner with higher isolation between the sig gen ports? Note that the match at the sum port of the combiner will usually dictate the actual level of isolation.
Does the device you are testing have a good match at the input? If not, then you may be getting less isolation between generators than you think. Sometimes you can add a small attenuator at the sum port to get the ideal compromise.
Also, have you checked to see if the IMD terms you are seeing are following the third order response when you change the drive levels? i.e. you may calculate a different IP3 for different drive levels and often devices that claim a high IP3 figure will quote the IP3 figure at a specific drive level. What (device output) test tone level are you currently using?