In a way, it depends on how decisions are made in the product category. For example, if it's a market for automobiles, maybe respondents could probably narrow down their choices to within a couple of categories (trucks, SUVs, sports cars, sedans, minivans). I don't imagine there are too many folks who intend to buy a red convertible but who are so taken at the lot with the sexy lines of the silver minivan that they jump so far across categories. Then you could build an experiment that traded off only the relevant product categories.
If you're modeling the liquor shelf at the grocery store or the cold remedies shelf at the pharmacy, however, respondents may not restrict themselves within categories quite so reliably and then getting an evoked set of brands (or SKUs) might make sense. The set of brands or SKUs I'd consider might cut across neat category boundaries.
On the other hand, even that might overly restrict your respondents' freedom in ways that will bias your understanding of their choices. I suppose folks might go to the store thinking about one cold remedy and then see other brands on the shelf and change their mind at the point of purchase. It could be that choices are malleable enough to be influenced in ways that using an evoked set might overly restrict. Maybe in this case you do want all the SKUs to come into and out of your choice experiment.
I wish I had a great answer for you that would apply at all times for all categories, but it's a tougher question than that!