Here is what probably is going on:
When respondents disagree about the order of preference for levels and yet you are creating a pooled model that averages across people, the computed scores by taking the range across levels within a conjoint attribute for average groups of people become compressed.
For example, imagine that 50% of your respondents prefer level 1 to level 2, but the other 50% prefer level 2 to level 1. On average, the scores will be tied and the "importance score" will appear to be zero. But, latent class could correctly identify that there are two separate groups, will compute the scores for the separate groups, and if you compute the "importance score" for the two groups separately and then afterward combine them, you will see that the importance score has grown considerably and is more accurate to what respondents actually are telling you.
For measuring the importance of attributes, individual-level (HB) estimation is probably best, where you are estimating importance scores for each respondent separately, normalizing them to sum to 100%, and then averaging them across people. Another thing to consider is whether the attributes have known and expected preference order across the levels. If you a priori believe that every rational person should believe that level 1 should be preferred to level 2, you could consider setting the importance of an attribute to zero if the respondent's data show the utilities in the wrong order. We have called that "strict importance" in the past. But, you could accomplish a similar thing by constraining the utilities during HB utility estimation to follow the expected a priori order of preference.