They don't. You are confusing product utility with attribute importance. A product's utility is found by summing the part worth utilities for one level from each attribute. Within each attribute they sum to zero, but across attributes using one level of each they usually don't sum to zero.
Importance scores, on the other hand, are measured by looking at the range of part-worths across the levels of an attribute. Though they sum to zero, the range (max minus min) will be non-zero even for binary attributes. The ranges can then be compared and percentages computed across attributes to find an "importance" for each attribute.