Satellite peaks in photoemission spectra are sensitive probes of electron correlation in materials and molecules. Leveraging these measurements for device design calls for theoretical models connecting them to physical and chemical guiding principles. However, satellites require a multireference description, seemingly precluding interpretable approximations. Here, we address this challenge by exploring the ghost Gutzwiller embedding framework for spectral satellite prediction, in which correlated systems are represented through a non-interacting quasiparticle model in an enlarged Hilbert space, turning satellite modelling into an interpretable band structure theory. We determine the quality of ghost Gutzwiller satellites with virtually exact benchmarks of small molecules, and combine it with coupled-cluster solvers to enable embedding larger fragments, bringing the method from toy models towards the realm of real molecules.