Counts averages across people. The utilities estimated from logit, latent class, or HB should be fairly similar to what you are seeing from counts (after you exponentiate the utilities...or alternatively after you take the natural log of the counts proportions).
But, importances from HB are estimated at the individual level. This can be very different (and more correct) than importances computed from aggregate tables of utilities (across people and segments) or importances estimated from aggregate counts.
If you have an attribute like brand or color where people disagree strongly about their preferences, on average it tends to damp and wash out this attribute to make it appear like it is less important than the individuals and segments behind the averages actually think.