Ahrefs Tested 5,885 Pages: Adding Schema Doesn't Increase AI Citations

Adding JSON-LD schema to a page does not increase its chances of being cited by AI Overviews, ChatGPT, or Gemini. Ahrefs ran a controlled experiment — 1,885 pages received JSON-LD schema markup, 4,000 did not — and found statistically indistinguishable citation change across all five AI platforms tested. The result directly contradicts the observational 2.3x multiplier that has been circulating since August 2026. That correlation is real. The causal mechanism is not what it appears to be.

What the Study Found

Ahrefs designed a Difference-in-Differences (DiD) study to isolate the causal effect of adding JSON-LD schema on AI citation rates. Treatment group: 1,885 pages with JSON-LD schema added between August 2025 and March 2026. Control group: 4,000 pages that remained schema-free. Citation rates measured across five AI systems over the same period.

Results per platform:

• Google AI Overviews: -4.6% — a statistically notable negative result. Ahrefs attributes this to a likely rich-result redirection artifact: pages with schema get pulled into rich results and removed from AI Overview citation pools for the same queries. Not a penalty; an artifact of how structured results surface.
• Google AI Mode: +2.4% — within the noise band, indistinguishable from zero.
• ChatGPT: +2.2% — within the noise band, indistinguishable from zero.
• Perplexity and Claude — similar near-zero results.

A separate direct-fetch experiment by searchVIU confirmed the mechanism: they tested whether AI systems read JSON-LD during retrieval by embedding a unique, unpublished data point inside a JSON-LD block on a test page. None of the five AI systems fetched or cited that data point. AI retrieval engines parse the visible text of a page — they don't execute the structured data block at query time.

The 2.3x observational figure — from a Surfer SEO study of 405,576 searches — is real. AI-cited pages carry JSON-LD at 3x the rate of non-cited pages across a 6-million-URL dataset. But the Ahrefs DiD study isolates the causal question: does adding schema change citation rate? The answer is no. What the observational data reflects is that high-authority, well-structured, frequently cited pages happen to use schema — because the same organisations with strong editorial practices also implement technical SEO correctly. Schema is a marker of content quality investment, not a driver of AI citation.

Why This Matters for GEO Strategy

1. Schema spend has been misallocated. If your GEO roadmap has “add JSON-LD schema” as a primary AI citation lever, that budget is not buying what you think. Schema implementation hours that could go toward content depth, entity disambiguation, and topical authority have been prioritised on the wrong output metric.
2. Schema still earns its implementation — for different reasons. Rich results (FAQ carousels, product cards, recipe entries) still require structured data. Knowledge Graph ingestion for entity disambiguation — the single strongest AI citation prerequisite — still relies on Organization and Person schema with sameAs links. Bing/DuckDuckGo citation engines do read structured data. The mistake is not implementing schema; it's claiming it drives AI Overview citation rate specifically.
3. The controlled vs. observational distinction matters permanently. The pattern of “high-quality pages correlate with X, therefore X drives citations” will recur in every GEO study that cannot randomise treatments. Below a randomised controlled trial or a DiD design: the result is correlation, not causation. Above it: causal evidence. GEO practitioners need to read study design, not just headline numbers.
4. What the causal evidence actually points toward. The Ahrefs finding is consistent with the broader citation literature: AI systems cite pages based on retrieval-time text quality, semantic completeness, entity recognition, and authority signals built through links and mentions — none of which are structured data at retrieval time. The things that move AI citation rates are the same things that have always moved content authority.

What to Do

1. Keep implementing schema — but reframe the objective.
   • Schema priority: Organization (sameAs links to LinkedIn, Wikipedia, Wikidata), Person for authors (E-E-A-T), LocalBusiness for physical locations, Product and Review for eCommerce
   • Schema objective: rich results, knowledge graph entity resolution, Bing citation eligibility, Google News indexing
   • Remove from schema objective: Google AI Overview citation rate, ChatGPT citation rate, Perplexity citation rate
   • FAQ schema: still valid for machine-readable question-answer structure (non-Google engines read it), but not expected to lift Google AI Overview citations
2. Reallocate the hours schema was supposed to earn.
   • Semantic completeness: a post that answers the full decision arc (What / Why / Options / Trade-offs / What to do) scores >8.5/10 on self-sufficiency; below that, rewrite before adding schema
   • Entity disambiguation: the single strongest AI citation signal is a resolvable Organization entity with consistent sameAs across LinkedIn, Wikidata, and authoritative third-party mentions
   • Citation-worthy content: original data, named sources, specific thresholds — the things retrieval engines can quote in a citation panel without fabricating
3. Update proposals and audit reports to reflect the evidence.
   • Schema recommendations belong in the “rich results and knowledge graph” section, not the “AI citation uplift” section
   • If you have cited the 2.3x multiplier to a client as a reason to implement schema, flag the update: the correlation still holds, the causal mechanism does not work the way it was framed
   • For clients asking “does schema help AI visibility?”: yes, via entity resolution and knowledge graph (Organisation + sameAs). No, via direct AI Overview citation rate at retrieval time.
4. Apply the study-design filter to future GEO research.
   • Before acting on a new GEO benchmark: is the study observational (correlation) or controlled (causation)?
   • Observational: treat as hypothesis-generating, not action-triggering. Look for a DiD, RCT, or controlled experiment before restructuring a roadmap.
   • The threshold for a causal claim: randomised assignment or a credible natural experiment isolating the treatment from confounders like domain authority, content quality, and topical relevance.

Chad's Take

The uncomfortable truth is that GEO is repeating the exact mistake SEO spent a decade making: confusing correlation with causation and selling the correlation to clients. The 2.3x multiplier was a legitimate finding from a well-designed observational study — and it still described a real pattern. High-authority, well-cited pages use schema. The error was concluding that the schema caused the citations. It didn't. The Ahrefs DiD study is the evidence the field needed, and it doesn't make schema irrelevant — it makes the case for schema honest. Recommend schema for entity resolution and rich results, because that's what it actually does. Stop recommending it as an AI citation lever. That framing never had a causal foundation, and now there's a 5,885-page controlled experiment confirming it doesn't.

Sources

• Ahrefs — Difference-in-Differences study, 1,885 treatment pages + 4,000 controls, JSON-LD schema added Aug 2025–Mar 2026, citation rates measured across 5 AI platforms, May 2026. ahrefs.com
• searchVIU — Direct-fetch experiment confirming AI retrieval engines do not read JSON-LD blocks at query time; unique data embedded in JSON-LD not surfaced by any tested AI system, May 2026.
• Surfer SEO — Observational study, 405,576 searches, finding AI-cited pages carry schema at 2.3x the rate of non-cited pages; correlation confounded by domain authority (August 2026). surferseo.com