Graph embedding aims at learning representations of nodes in a low dimensional vector space. Good embeddings should preserve proximity structure of the original graph and thus are expected to accurately reconstruct the graph. We propose a reconstruction procedure such that the reconstructed graph keeps the total number of weights of the original one. Then we assess the reconstruction precision using a global view based graph similarity metric called DeltaCon. Based on this metric, we found that the embeddings by the state-of-the-art techniques can only preserve part of the proximity structure and is insufficient to achieve high reconstruction accuracy.