At Sawtooth Software, we offer rules of thumb to researchers to help them think about reasonable setups for their CBC questionnaires. However, each study is different. How many attributes the respondent must trade off, the length of the text involved with the levels, respondent heterogeneity, how many concepts are shown per task, the subject matter and engagement of the respondents. These all matter to accurate HB estimation for CBC studies.
Of course, with HB estimation, we're concerned that we have adequate information to estimate each attribute level. So, rules of thumb based on aggregate logit (pooled logit) don't seem to address your question.
John Howell, formerly of Sawtooth Software, provided the rule of thumb for CBC that each attribute level should appear about 6 times per respondent. It seems like a useful rule of thumb, but I haven't seen it formally tested. As far as my memory serves, I think the only place we've written this 6x rule of thumb is in our 2- and 3-day CBC training workshops. That's a PowerPoint deck.
In terms of empirical evidence, Hoogerbrugge and van der Wagt did a meta study of 35 CBC studies, examining the hit rate to holdout concepts as the tasks increase. That paper is entitled, “How Many Choice Tasks Should We Ask?” Hoogerbrugge and Van Der Wagt, 2006 Sawtooth Software Conference. You may download it from:
However, their study did not examine how many times each level for complex attributes appeared.