In wireless mesh routing, it is common to use periodic estimation of link quality to identify high throughput paths. While extensive work has been devoted to the design and static throughput evaluation of path metrics, evaluating the accuracy of the underlying link cost metrics under dynamic conditions has not received due attention. We introduce an experimental and analytical methodology that quantifies the ability of link cost metrics to accurately estimate link capacity. We use this methodology on a wireless mesh testbed to evaluate network layer and cross-layer estimation approaches of ETT, a popular wireless link cost metric. Our results show that the network layer approach exhibits low correlation with link capacity across time and across multiple links. On the other hand, cross layer information obtained by our methodology can significantly improve accuracy and can aid in identifying the dominant factors that lead to link cost metric inaccuracies.