Predictive vision–language learning aims to learn semantic representations by predicting latent targets rather than reconstructing pixels or generating text. We ask whether latent prediction alone induces the global image–text geometry required for zero-shot matching, and how the location of alignment pressure changes this outcome. In a controlled joint-embedding predictive model family with matched training and evaluation protocols, non-contrastive latent prediction retains information recoverable by supervised adapters and token-level probes but does not form a usable zero-shot image–text space under pooled cosine scoring. A target-contrastive design improves matching, but remains weaker than the directly aligned variant. Our results show that latent prediction is underconstrained at the global embedding level in this setting. Direct global alignment supplies the constraint that makes predictive representations usable for zero-shot scoring. With only a modest direct global alignment loss, latent prediction becomes competitive with matched CLIP and SigLIP baselines and improves several language-sensitive diagnostics.