The precise computation of abstractions is a bottleneck in many approaches to CEGAR-based verification. In this paper, we propose a novel approach, based on the use of structural information. Rather than computing the abstraction as a single, monolithic quantification, we provide a \emph{structure-aware} abstraction algorithm, based on two complementary steps. The first, high-level step exploits the structure of the system, and partitions the abstraction problem into the combination of several smaller abstraction problems. This is represented as a formula with quantifiers. The second, low-level step exploits the structure of the formula, in particular the occurrence of variables within the quantifiers, and applies a set of low-level rewriting rules aiming at further reducing the scope of quantifiers. We experimentally evaluate the approach on a substantial set of benchmarks, and show significant speed ups compared to monolithic abstraction algorithms.