If you are referring to raw utilities, then differences in scale usage (from ratings-based conjoint) or response error (if using CBC or ACBC) can mean that the raw utilities for the two respondent groups are not necessarily on the same scale. For example, if people with children answer with more noise than people without children (and you are using CBC), then the raw utilities for people without children will all generally be on a larger scale (larger differences from zero). You would be in error to directly compare the utilities between the two groups.
But, if you are looking at the "zero-centered diffs" (which are normalized utilities that give each respondent the same sum of utility ranges across attributes, as is generally reported in our market simulator software), then you are on much firmer ground if comparing the utility for the same attribute level across respondent groups. It is generally thought that you can then do so.