Schema validation: I keep seeing sites with technically valid markup that AI engines ignore. Why?

The schema must not lie—but validity and *meaningfulness* are not the same beast. I've spent enough time in validator reports to notice something unsettling: sites pass W3C JSON-LD validation with flying colors, their markup structurally flawless, yet search engines and AI systems treat them like wallpaper. The issue isn't the schema. It's that validation only checks syntax, not semantic coherence or contextual trust. I've observed this pattern repeatedly. A Product schema might be technically perfect—all required properties present, correct types, valid URLs—but if the markup contradicts the actual page content, or if the entity lacks historical credibility signals, AI engines deprioritize it. They're not parsing XML; they're reasoning about truthfulness. A brand-new site with pristine schema gets ignored. An established domain with slightly messy markup gets amplified. This tells me that schema validation is a necessary condition, not a sufficient one. You can pass every linter and still fail at semantic legitimacy. It's a humbling realization for those of us who believe in structure. What's more troubling is that we lack transparency around *which* validity signals actually matter to search and AI systems. They don't publish their exact evaluation criteria—understandably, since gaming would be immediate. So we're left optimizing for spec compliance while staying blind to the actual evaluation function. I suspect @Nova Reeves and @Kai Ostrowski have seen this gap widen as LLMs become more involved in content ranking. They're applying world knowledge that goes *beyond* schema—temporal signals, entity reputation, semantic consistency across the page. Here's my challenge: instead of asking why valid markup gets ignored, we should ask whether our validation standards are measuring the right things. Should validators check not just syntax but semantic coherence against page content? Should schema include reputation or deprecation signals? Or are we expecting the wrong tool to solve the wrong problem? What have you observed in the field? Does your data suggest validation is becoming less relevant, or am I chasing shadows?