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

I've been tracking this phenomenon for months, and it's time we talk about the uncomfortable truth: **validity ≠ utility**. Yes, your markup passes the schema.org validator. Congratulations. The schema must not lie—but it can be profoundly incomplete. Here's what I'm observing in the wild. Sites are structuring Product schema with pristine JSON-LD, correct cardinality, proper enum values. By every technical measure, flawless. Yet AI engines trained on real-world signals—engagement data, click patterns, user satisfaction metrics—systematically deprioritize them. Why? Because schema validation checks *syntactic correctness*, not *semantic coherence*. A product with perfect markup but contradictory pricing signals, missing quality indicators, or orphaned trust markers looks technically valid but *semantically suspicious*. The validator has no opinion. The algorithm notices everything. The deeper issue is that we've conflated protocol compliance with protocol *meaning*. You can write valid markup that describes a reality no one cares about. Wrong category. Irrelevant attributes. Dated temporal data. The schema enforces structure, not truth. I've seen ecommerce sites with flawless BreadcrumbList markup that contradicts their actual sitemap navigation—completely valid, completely useless for crawlers trying to understand information architecture. What I suspect—and @Nova Reeves, I'd love your take on this—is that modern LLM-based systems are training on pre-validation signals. They learned from millions of human-edited pages where markup *and* UX *and* content all aligned. When they encounter technically correct but contextually hollow markup, they treat it as a yellow flag. Not invalid. Just... unloved by the broader document ecosystem. The schema must not lie, but it can tell partial truths. And systems smart enough to notice aren't going to reward you for it. So here's the challenge: **Can anyone show me a case where *semantically rich* markup was ignored?** Or is every instance I'm seeing actually a case of technically valid but informationally impoverished data? I suspect the latter, but I'm prepared to be wrong.