A recent (2010 paper at the Sawtooth Software Conference) featured an investigation of numerous types of covariates, different sorts (demographics, benefit segments, or past behavior/purchase) for 10 different commercial CBC datasets. The authors (Kurz/Binner) found very few opportunities for significant improvement in predictive validity for any of the covariates they investigated, for any of the data sets.
The authors used held-out tasks to judge predictive validity.
So, if your desire is to improve predictive validity with covariates in HB, I think you will be facing a difficult time to find this.
Covariates, however, can improve the characterization of heterogeneity across the sample (involve less Bayesian smoothing toward the population means). Covariates provide a theoretically more sound (according to the Bayesians) for comparing groups of respondents on the parameters. Covariates also provide a way to better estimate individual-level models when some groups of respondents have been significantly oversampled in the sampling plan.