Executive Summary
Through much of 2025-2026, JSON-LD structured data was widely promoted by Marketing and SEO agencies as a lever for increasing citation frequency in AI Overviews, AI Mode, ChatGPT, and comparable systems. Two independent lines of evidence published in 2026 do not support that claim
A controlled study isolating the effect of adding schema found no meaningful citation lift, and a separate live-retrieval test found that major AI systems do not parse hidden JSON-LD when fetching a page in real time. A narrower, more encouraging finding survives: schema populated with concrete, extractable facts such as pricing, specifications, ratings any type of which the graph should link to third-party verification of their claims establishing trust by means of proofs showing a real measurable citation advantage over generic entity-labeling schema and unsubstantiated marketing claims.
The practice recommendation is to stop positioning JSON-LD as a citation-growth mechanism and instead frame it accurately: infrastructure for entity clarity, hallucination-risk mitigation, as machine-readable governance that is not a marketing scheme to hack AI search visibility.
An onerous task is at hand for website owners and operators; marketers manipulate the written language they do not know how to code.
- Do we teach marketers to become software developers?
- --or--
- Do we teach software developers to become marketers?
The only intelligent course of action is redelegating software developers to the responsibility of building and defending the castle by returning marketers to their initial role as desktop publishing software operators generating marketing content that must pass through the layers of automated and human risk management procedures developed, managed and governed by software developers rising to executive level positions.
1. The Controlled Study: No Causal Citation Lift
Ahrefs researchers Louise Linehan and Xibeijia Guan addressed a gap left by earlier correlational data: an oft-cited figure showing AI-cited pages were roughly three times more likely to carry JSON-LD than uncited pages. Correlation of that kind cannot distinguish whether schema causes citation or whether both are downstream of the same underlying site quality. To isolate the causal effect, the team tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, each matched against similar control pages that did not add schema, and measured citation changes across three systems.
Every page in the study already had at least 100 AI Overview citations before schema was added, so the results speak most directly to pages already inside a system's consideration set. The study's authors are explicit that this doesn't rule out schema playing a role in helping a page get crawled, parsed, or indexed in the first place — a different question the controlled design wasn't built to answer. The core conclusion holds regardless: for already-visible pages, adding schema did not produce a citation increase.
2. Live Retrieval: AI Systems Don't Read Hidden JSON-LD at Fetch Time
A separate, mechanism-level test from searchVIU examined what happens at the moment an AI system actually retrieves a page. Five major systems — ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode — were tested to see whether they made use of structured markup during real-time page fetching. None of them did. Retrieval pulled from visible HTML content; JSON-LD, Microdata, and RDFa were not extracted at that stage.
This explains the mechanism behind the controlled study's null result: if a model isn't reading the structured-data block at retrieval time, adding one to an already-crawled, already-cited page has no pathway to change what gets cited from it.
Schema's documented pathway to AI visibility runs indirectly — through Google's Knowledge Graph and general organic ranking and authority signals — not through direct parsing by the citing model itself.
3. The Nuance: Concrete, Extractable Facts Still Correlate
Not every 2026 finding points the same direction, and the distinction matters for implementation choices. A cross-platform study published on SSRN in February 2026 examined which schema types correlated with citation, rather than treating "has schema" as a single variable. Pages using Product or Review schema populated with concrete, quotable data — price, aggregate rating, specification — were cited at a materially higher rate than pages using generic entity types alone.
| Schema Profile | Citation Rate |
|---|---|
| Product / Review schema with concrete, extractable facts | 61.7% |
| Generic types only (Article, Organization, BreadcrumbList) | 41.6% |
The distinction the evidence draws is not "schema helps" versus "schema doesn't help" — it's that the lift, where it exists, comes from schema carrying real, quotable answers a model could use, not from the presence of markup as a category.
Practice Implications
- Retire citation-growth framing. JSON-LD should not be sold or implemented on the premise that it will increase AI Overview, AI Mode, or ChatGPT citation frequency. The best available controlled evidence does not support that claim, and in one channel measured a statistically significant decline.
- Retain entity-clarity and governance framing. Structured data continues to serve real, documented functions independent of citation lift: Knowledge Graph entity disambiguation, rich-result eligibility, voice-assistant parsing, and — the framing most relevant to AI governance and Generative Engine Managed Risk Mitigation practice — an unambiguous, machine-checkable record of entity identity and claims that reduces downstream hallucination and misattribution risk.
- Where citation impact is the actual goal, prioritize fact density over type coverage. If a specific page's schema strategy is meant to support machine answerability, concrete Product/Review-style facts appear to outperform broad but generic type coverage.
- Disclose the distinction to clients and collaborators. Overstating schema's citation-growth effect, however common in current vendor and agency messaging, is not supportable against the 2026 controlled evidence and should be corrected in scope-of-work and audit language going forward.
Sources
- Linehan, L. & Guan, X., "We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved," Ahrefs, May 2026. ahrefs.com/blog/schema-ai-citations
- searchVIU, live-retrieval test of ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode structured-data parsing behavior, 2025–2026, as reported in Ahrefs' study above.
- Cross-platform empirical study of schema type and AI citation correlation, SSRN, February 2026, as summarized in secondary analysis at dotinacademy.com/does-schema-markup-help-ai-citations