It depends on how the replication is done, but the big replication projects typically use very large sample sizes compared to the original studies, so their error bars are much smaller.