Thesis
Over the past several years, structured data, JSON-LD, and Generative Engine Optimization strategy have largely been claimed as marketing territory — sold, scoped, and implemented by marketing agencies and in-house marketing teams, often with limited source-code literacy. The 2026 controlled evidence on schema and AI citation behavior, examined in a companion report, shows that many of the technical performance claims marketing has made about this work — principally, that adding schema increases AI citation — do not hold up under causal testing. That gap between claim and evidence is not incidental. It is a predictable consequence of technical infrastructure being owned by a discipline whose training and incentives are oriented toward messaging and campaign performance, not toward code correctness, graph integrity, or falsifiable claims about how machines actually retrieve and parse content.
The practical recommendation that follows: implementation, validation, and testing of machine-readable infrastructure — structured data, entity graphs, AI-retrieval verification — should sit with developers and other code-literate practitioners who can build it correctly and test whether it works. Marketing's role should return to what it does well and can defend with evidence: message content, brand voice, campaign strategy, and creative production. In the language of an earlier era of desktop computing, marketing's proper domain is publishing — producing and curating the words and creative that appear on a page — not owning or making unverified claims about the machine-readable infrastructure beneath it.
1. The Evidence Behind the Claim Gap
The clearest documentation of this gap comes from a 2026 controlled study that isolated the causal effect of adding JSON-LD schema to pages already visible to AI systems: no meaningful citation lift was found across three major platforms, with one channel showing a statistically significant decline. A separate live-retrieval test of five AI systems found that none of them parse hidden structured data at the moment of fetching a page — they read visible content. Yet the claim that schema drives AI citation has been, and in many quarters continues to be, standard marketing and agency messaging.
This is not evidence that every marketing claim about GEO is false. It is evidence that a significant, widely repeated claim went unverified for years in a discipline that markets technical capability without consistently testing it — and that the testing, when it finally happened, required a controlled-study methodology and retrieval-level technical understanding that sits outside conventional marketing practice.
2. The Skills Gap Underneath the Claims Gap
Correctly implementing and validating machine-readable infrastructure requires a specific skill set: understanding JSON-LD and RDF graph semantics, recognizing how server-side rendering differs from client-rendered content a crawler may never see, using validation and graph-expansion tooling to catch structural errors invisible to surface-level checkers, and reading a site's actual served HTML rather than assuming a CMS plugin did the job correctly. None of this is inherently mysterious, but it is a code-literate skill set, not a copywriting or campaign-management one.
The practical consequence shows up directly in the kind of defect this report's companion engagement uncovered: a duplicate entity identifier silently merging conflicting business data, a dangling reference pointing at a class definition instead of a real node, a non-standard property invented for convenience that validators don't flag but that never actually resolves to anything a machine can use. Surface-level validators — the tools most marketing-run implementations rely on exclusively — do not catch this class of error. Catching it required expanding the graph with a JSON-LD processor and reading the resulting RDF structure directly, which is development work, not marketing work.
3. Why This Is a Risk-Management Issue, Not Only a Marketing One
For statutory licensees — architects, engineers, and other regulated professionals — the stakes of this skills gap go beyond wasted marketing spend. Structured data published under a licensee's name is a machine-readable representation of their professional identity and claims. An error a marketing team cannot see — a credential misattributed, a service area overstated, an entity reference that silently resolves to the wrong node — becomes a standing, difficult-to-detect misrepresentation that AI systems and search engines may propagate on the licensee's behalf, without the licensee ever knowing it's wrong.
This reframes the ownership question. The people responsible for a licensee's structured data should be capable of reading what they published, verifying it resolves correctly, and testing what it actually causes to happen — capabilities that, by the evidence above, correlate with development skill more than with marketing training.
4. A Proposed Division of Responsibility
| Function | Owned By | Why |
|---|---|---|
| Structured data implementation, validation, and testing | Developers / code-literate technical practitioners | Requires graph-level reasoning and tooling most validators and marketing training don't provide |
| Claims made to clients about technical performance | Whoever can produce the controlled evidence for the claim | The 2026 findings show unverified performance claims are the recurring failure mode |
| Message content, brand voice, campaign strategy, creative production | Marketing | Genuine marketing expertise the technical discipline does not typically hold |
5. A Fair Accounting of the Counter-Case
Where this thesis overreaches
Framing this as marketing broadly having "usurped" the website understates real marketing expertise that development does not typically supply: audience research, brand narrative, customer-journey design, and UX writing are disciplines in their own right, not lesser tasks that reduce to formatting text. A strict hand-back-the-keys model risks discarding that expertise and recreating the opposite failure — technically correct infrastructure serving content that answers no real audience need.
It's also worth separating two different failures the evidence actually documents. The controlled study shows unverified performance claims about schema were wrong. It does not show that marketing professionals as a discipline are broadly responsible for the wider erosion of trust and quality across the web — that erosion has many more proximate causes (ad-supported incentive structures, platform algorithm changes, low-quality content farms, actual malicious activity) that are not specific to structured-data marketing practice and shouldn't be folded into this argument as if the evidence established that link. The defensible claim is narrower and stronger: a specific, widely repeated technical claim went untested for years, and the people making it lacked the tooling to test it themselves. That's sufficient to support returning technical ownership without needing a broader indictment the evidence doesn't back.
Most organizations that get this right in practice don't fully separate the roles — they build collaborative models where developers own correctness and marketers own message, with a documented interface between the two rather than a wall. That interface, not a full role reversal, may be the more durable long-term structure.
Recommendation
- Structured data and other machine-readable infrastructure should be implemented and validated by practitioners who can read, expand, and test the graph directly — not solely by surface-level validator pass/fail status.
- Any technical performance claim made to a client or the public (citation lift, ranking impact, AI visibility) should be sourced to controlled evidence, not repeated because it is common industry messaging.
- Marketing's ownership of message content, brand strategy, and creative production should be preserved and respected — the goal is a documented division of responsibility, not a demotion of the discipline.
- Audit and disclosure practices (consistent with existing engagement protocols) should apply equally to technical and marketing claims: state what was actually verified, and by what method, rather than what is generally believed to be true.