A Perfect Schema Score That the AI Ignored

Structured data has become a reassuring green checkmark. Run the validator, watch every type resolve, confirm the rich-result eligibility, and move on satisfied. The implicit promise is that valid schema is a lever on machine understanding — that if the syntax is clean, the machines will read, trust, and surface the page accordingly.

It is a half-truth, and the missing half is expensive. A page can earn a flawless schema score and still be ignored by the AI answer it was supposed to win. Syntax validity confirms that the markup is well-formed. It says nothing about whether the markup describes the things that actually matter, or whether the claims it asserts are corroborated by the visible page. The validator checks grammar; the AI checks meaning.

Coverage is not the same as correctness

There are two questions a structured-data audit should answer, and most stop after the first. The first is syntactic: is the markup valid? The second is semantic: does the markup actually cover the entities, attributes, and relationships a model needs to answer the questions buyers ask? A site can score perfectly on the first while leaving the second almost empty — valid Organization and WebPage markup wrapped around pages whose real subject matter is never described in a machine-readable way at all.

This is the trap of the green checkmark. Coverage measures how much of what matters is described; syntax measures only whether what is described is well-formed. A page can be 100% valid and 20% covered, and the validator will congratulate you on the 100% while the AI quietly works from the 20%.

Why models discount unsupported markup

Even semantically rich markup is not automatically trusted. AI surfaces are wary of structured data as a self-serving signal — markup is, after all, a claim the publisher makes about itself. When the JSON-LD asserts a fact the visible page never states, or states differently, the model has every reason to discount the markup and reach for a source whose claims are corroborated on the page a human would read.

The lesson is that schema is a grounding aid, not a grounding substitute. Its job is to make explicit and unambiguous what the page already demonstrably says. Markup that races ahead of the visible content — asserting ratings, attributes, or relationships the page does not support — is not a head start; it is a credibility liability the model is built to detect.

What a useful structured-data audit looks like

A structured-data audit worth running reports two numbers, not one. Syntactic validity tells you the markup will not break. Semantic coverage tells you how much of the answerable surface area — the entities and attributes behind the questions you want to win — is actually described and corroborated. The gap between those two numbers is the work.

Closing it is rarely about adding more schema types. More often it is about pointing the markup you already have at the things that matter, and making sure every machine-readable claim is matched by a human-readable one on the page. The same evidence that earns a citation from the AI answer is the evidence that justifies the markup — which is why the audit that exposes the schema gap and the audit that exposes the citation gap keep turning out to be the same audit.