Learning World models from offline trajectories enables agents to accomplish different tasks through planning. Using object-centric (OC) representations that decompose the scene into slot representations that bind to scene objects has been proposed as an inductive bias for learning sample-efficient world models that generalize. Yet prior object-centric world models (OCWMs) take the object-centric slot encoder as given and evaluate in-distribution, leaving open whether the object-centric bias actually delivers for planning and what within the OCWM drives it. We conduct a controlled study of OCWMs for visual model-predictive control along two axes: object-centric representation quality and the generalization induced by object-centric representations versus scene-centric representations. We find that (i) planning success correlates strongly with unsupervised slot-quality metrics; (ii) that with well-bound slots, we can remove reliance on redundant or unnecessary proprioceptive inputs and masking inductive biases that prior methods relied on; and (iii) a well-trained OCWM plans more robustly than scene-centric world models to unseen environment distribution shifts.