That leads to a much-noted chicken-and-egg problem... but that aside, for all but the most menial and interchangeable X, employers don't generally have access to data about how well and how long prospective hires have done X. They have access to candidates' word for how well they've done more or less imperfectly related work, and usually to recommendations from their former employers and coworkers -- but the former is unreliable, and the latter demonstrates only that the candidate isn't a complete schlub.

I haven't read the paper in the ancestor, but it seems reasonable to me that IQ would often end up being a better predictor of performance, given these constraints.