the neat thing about conjoint is that it multiplexes the choice-making (when designed well), so that the choice burden scales reasonably while also better representing collective choice. nothing is going to be perfect, but conjoint could potentially be better than all current systems, probably by an order of magnitude or more.
it's relatively complicated, so yes, 'experts' (i disdain this term, frankly) could in theory manipulate the system, but transparency of voting data (just like now) and survey design (the more likely bias vector) would make that difficult. transparency would allow competing independent journalists, academics/researchers, political analysts, etc. to weigh in and balance out biases (not perfectly of course), as well as verify results.
Additionally, all of this ignores the possibility that a representative could change their stance on an issue after receiving more information on an issue. Presidents and Supreme Court justices have changes their stances once their terms have begun or events have taken place during their time in office. I would hate a system that relies on tightly binding the representatives to narrow definitions of hot-button issue stances.
For politicians, for me, it's more about (perceived) character vs. their stated policies.
currently, there is no option to choose a candidate supporting marijuana legalization, $15 minimum wage, pro-life, and pro-gun policies (as an example), since we practically only have two relatively fixed packages of pre-ordained, all-or-nothing policy positions to choose from. revealed preferences would pressure parties into fielding candidates that better match the non-partisan preferences of constituents. or better yet, it would lead to the waning of the two-party system altogether, toward a more direct representative democracy.
one method of implementing this (from thinking about it for all of 2 minutes), is developing a list of the top 10-20 issues that voters care about in a given election and providing tradeoff choices for each, then matching them to the candidate closest to their revealed preference. of course, the biggest issue with method is that it moves the problem up one level, from voting to choice design.
[0]: https://wikipedia.org/wiki/Conjoint_analysis