Generative Adversarial Networks (GANs) are now able to generate astonishingly realistic high-resolution images. Recent work has shown the emergence of semanticallymeaningful manipulations simply by editing the corresponding latent vector. However, a real image must first be inverted into its GAN latent code before editing. Previous work usually achieves accurate reconstruction, but poor-quality latent vectors: applying known editing methods onto these latent codes results in artifacts and erroneous edits. We aim to bridge the gap between reconstruction and editability. We propose a novel instance-optimization based inversion method, which specifically aims to maximize the semantic information of the latent vector, all while producing an accurate reconstruction. We introduce the iMAGe-latEnt Consistency loss ("MAGEC"), which allows supervision in the latent space, encouraging editability of the resulting latent vector. We provide extensive qualitative and quantitative evaluation to validate our method, using the recent state-of-the-art StyleGAN and show that our method outperforms baseline inversion methods, opening the door to new realms of real-image editing.