Structured Data and Machine Trust: What the Evidence Shows vs. What Vendors Claim

What structured data actually is: a standards-body vocabulary, not an SEO trick

Structured data is a controlled vocabulary for stating, in a form a machine can parse without inference, what a page is and how its facts connect. The dominant implementation is schema.org, expressed most often as JSON-LD embedded in the page. Its provenance is documented and unusually collaborative for a web standard: schema.org was founded in 2011 as a joint effort of Google, Bing, Yahoo, and Yandex, competitors agreeing on one vocabulary precisely so that machine-readable markup would mean the same thing across every engine that consumed it.

That lineage matters because it locates structured data in a tradition much older than the current AI-search cycle. Schema.org sits on top of the Linked Data program articulated by Bizer, Heath, and Berners-Lee, whose principle was that the web should expose entities and relationships that machines can follow, not merely documents for a person to read. Google operationalized the same idea publicly in 2012 with the Knowledge Graph, launched with roughly 500 million entities and 3.5 billion facts under the explicit thesis of indexing "things, not strings."

The upshot: structured data is a governed, standardized, long-lived layer of web infrastructure, not a proprietary growth hack that a vendor discovered. Treating it as a durable technical foundation is well supported. Treating it as a lever with a fixed conversion rate into visibility is not the same claim, and it is not supported by the same evidence.

What the documented mechanics deliver, and how "does schema markup improve rankings" actually gets answered

Google's own structured-data documentation is precise about what the markup does. It recommends JSON-LD as the format, instructs publishers to mark up only content that is visible to users, and states that valid structured data makes a page eligible for rich results, the review stars, FAQ expanders, and enhanced listings, without guaranteeing that any rich result will display. Eligibility is controllable; display is Google's decision at query time.

Two things follow that most commercial content blurs. First, structured data is a parsing and eligibility mechanism, not a documented ranking booster in its own right. Google has repeatedly stated that most structured-data types are not ranking factors; the benefit is that a correctly marked-up page can appear in a richer form and be read unambiguously, not that its position rises because markup is present. Second, the answer to "does schema markup improve rankings" is that it changes how a page can be presented and parsed, and can improve click behavior where a rich result displays, but the standards documentation does not promise a ranking movement, and neither should anyone citing it.

These are the well-evidenced mechanics: standardized vocabulary, unambiguous machine parsing, and eligibility for specific rich features that a page genuinely qualifies for. Everything in this paragraph traces to a primary source you can open and check. That is the boundary of what the documentation establishes.

Where the vendor claims outrun the evidence

The gap opens when documented mechanics are restated as quantified outcomes. A recurring genre of industry-blog claim attaches a precise number to structured data, an "accuracy jump," a percentage lift in visibility, a fixed increase in rich-result click-through, presented as though it were a measured law. The problem is not that structured data is worthless; it is that these figures rarely cite a controlled study, are frequently vendor analyses of their own client base with no counterfactual, and travel from post to post detached from any method that would let a reader verify them.

There is a well-documented precedent for this pattern in the same domain. E-E-A-T, the Experience, Expertise, Authoritativeness, and Trust framework, is described in Google's own Search Quality Rater Guidelines as a heuristic for human raters evaluating the algorithm's output, and Google has explicitly stated it is not a scored ranking factor a page can optimize into a document. Yet a large share of commercial advice treats E-E-A-T as a machine-readable dial. Structured data attracts the same category error: a real, documented mechanism gets recast as a quantified optimization target it was never shown to be.

The rigorous posture is not scepticism about structured data. It is scepticism about any number attached to it that does not come with its working. When a claim says markup produced a specific lift, the questions are: measured against what baseline, over what population, with what control, and published where. If those answers are missing, the number is an assertion wearing the costume of a finding.

Schema markup for AI search: what the peer-reviewed evidence does and does not say

The most consequential over-claim is that structured data is how you get into AI answers. Here the primary sources are unusually clear, and they cut against the marketing. Google's guidance on optimizing for AI features states plainly that those features run on its core index and ranking systems and that no special markup is required to appear in AI Overviews. Structured data is not the admission ticket to a generative answer.

The founding peer-reviewed study of the field points the same way. The GEO paper, published at ACM SIGKDD in 2024, ran a controlled benchmark of roughly 10,000 queries and measured which content interventions changed whether a source was surfaced in a generated answer. The levers that moved visibility by 30 to 40 percent were content-level: adding citations to credible sources, including direct quotations, and replacing vague claims with specific statistics. Citing authoritative sources was the single strongest lever. Structured-data markup was not identified as the mechanism that moved AI visibility in that study.

A separate, sobering caveat sits underneath the whole idea of machine "trust." Even when a generative system does cite sources, the citation and the reasoning are frequently decoupled: an attribution survey of production retrieval-augmented systems reported that over 95 percent of answers from tested open-source models contained at least one unattributed sentence, and other work shows citations can be attached to an answer independently of the evidence that actually produced it. Being marked up, retrieved, and even cited is not the same as being read faithfully. That is a reason to state plainly how much control markup confers, not to oversell it.

The entity-identity caveat: why "sameAs" is looser than it sounds

Structured data's most ambitious promise is entity resolution: declaring, through properties like sameAs, that the business on this page is the same real-world entity as the one on that profile, so every engine converges on one identity. The mechanic is real and worth engineering. The strength of the guarantee is routinely overstated.

The foundational Semantic Web analysis of exactly this mechanism, Halpin and colleagues' study of owl:sameAs, the identity primitive that schema.org's sameAs inherits from, found that publishers across the Linked Data web use it inconsistently, applying at least four looser, non-equivalent senses of "same." The underlying looseness has not been fixed; it has been worked around by platform-specific heuristics. So "we made your business the same entity everywhere" is a defensible engineering goal, but it is a probabilistic reconciliation that engines still adjudicate, not a hard identity a single markup field can assert into existence.

How to read a structured-data claim

The practical test is to sort every statement about structured data into one of two bins before acting on it: a documented mechanic you can verify against a primary source, or a quantified outcome that needs a study behind it. The first bin is where structured data earns its place in a technical foundation. The second is where most of the hype lives.

• Documented and verifiable: JSON-LD is the recommended format; mark up only content visible on the page; valid markup makes a page eligible for specific rich results; correct markup lets engines parse your facts unambiguously. Source: Google's structured-data documentation and schema.org.
• Documented and verifiable: no special markup is required to appear in Google's AI features. Source: Google's AI-features optimization guidance.
• Established by peer-reviewed study: the content-level levers that moved AI-answer visibility were citations, quotations, and statistics, not schema markup. Source: the GEO paper, SIGKDD 2024.
• Needs a primary source before you believe it: any specific percentage "accuracy jump," "ranking lift," or "AI citation increase" attributed to adding schema. Ask for the baseline, the control, the population, and the publication. Absent those, treat it as an unverified vendor claim.
• Overstated by default: "we will make you the same entity everywhere" as a hard guarantee. The identity mechanism is real but probabilistic, and engines still adjudicate it.

What this means for a technical foundation

Read this way, structured data comes out looking better, not worse. Stripped of the inflated numbers, it is one of the few parts of the discovery stack whose value is inspectable end to end: an open standard, public validators, and eligibility you can confirm on your own live pages. That verifiability is precisely why it belongs in a technical foundation and precisely why it should never be sold with a fabricated lift attached.

The correct engineering standard is narrow and accurate. Mark up only what is true and visible. Choose the schema type that fits the page. Resolve the entity graph deliberately, knowing the identity claim is a reconciliation rather than a decree. Verify every block in Google's validators, and report eligibility, which is controllable, rather than promising a rich result or an AI citation, which is not. Done to that standard, structured data is a durable asset. Dressed up with an unsourced percentage, it becomes one more claim a discerning buyer should discount.