Please note that it is difficult to say something generic about all VRP variants. There are certainly a lot of counter examples of what I wrote below.
First, column generation is mainly used within branch-and-cut-and-price algorithms which are the best exact methods for most VRP variants, even if branch-and-cut algorithms remain better in a few cases. You can give a look at the VRPSolver package for example.
Given an infinite time, branch-and-cut-and-price algorithms will find the optimal solution of a VRP problem and prove that it is optimal. Whereas local search based metaheuristics may find it, but won't prove that it is optimal.
Unfortunately, when branch-and-cut-and-price algorithms fail to terminate within the given time limit, they usually don't terminate with a solution as good as the solution found by local search based metaheuristics within the same time limit.
A rule of thumb would be, up to 200 locations to visit, branch-and-price-and-cut algorithms should find the optimal solution quickly; above, local search based metaheuristics would yield better solutions for running times of the order of 1 hour. Of course, this depends on the variant and the dataset.
In addition, column generation can also be used in a heuristic fashion, which can yield better solutions than a whole branch-and-cut-and-price, and quicker. If I'm not mistaken, this is for example the approach chosen by the VRPy package.
Something to look at to determine if it might work better than local search based metaheuristics is how much the problem is constrained. More constraints will make local search less effective, but the pricing problem of the column generation easier. Still, in the literature on VRP problems, local search based metaheuristics are far more present that column generation based heuristics.