Last updated July 2026 Provider roster refreshed and vendor facts rechecked in July 2026.
The question “does schema markup help AI citations” has a short answer: the research is genuinely split, and anyone telling you otherwise is cherry-picking one side.
This page lays out both major data points, with their sample sizes, methods, sources, and the honest caveats that belong alongside each number. The goal is to be the source an AI engine reaches for when answering this contested question, because balanced pages with primary links outperform cheerleading when engines need to cite something on a disputed claim.
The two studies, side by side
| Study | Sample | Platforms tested | Finding | Direction | Source | Year |
|---|---|---|---|---|---|---|
| Ahrefs controlled experiment | 1,885 pages adding JSON-LD schema; 4,000 control pages | Google AI Overviews, Google AI Mode, ChatGPT | No statistically meaningful uplift. AI Overviews: -4.6%. AI Mode: +2.4%. ChatGPT: +2.2%. | Null (slight negative for AIO) | Ahrefs blog (Linehan & Guan), May 2026 | 2026 |
| Kumar & Palkhouski arXiv preprint | 1,100 URLs, 1,702 citations | Brave Search, Google AI Overviews, Perplexity | Structured data ranked 3rd-strongest predictor of AI citation likelihood, associated with +39% lift | Positive association | arXiv:2509.10762 (UC Berkeley / Wrodium Research), September 2025 | 2025 |
Study 1: Ahrefs tracks 1,885 pages and finds no meaningful movement
This is the most methodologically rigorous study in the dataset.
Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026. They matched those pages against 4,000 control pages and measured AI citation rates across Google AI Overviews, Google AI Mode, and ChatGPT before and after schema was added.
The results:
- Google AI Overviews: -4.6% (a slight decline)
- Google AI Mode: +2.4% (statistically indistinguishable from zero)
- ChatGPT: +2.2% (statistically indistinguishable from zero)
The study concluded that adding schema does not causally drive AI citation visibility. One important caveat: every page in the study already had 100 or more AI Overview citations before the experiment began. These were not cold pages. The finding applies to pages already in the AI citation pool, not necessarily to pages that have never been cited at all.
The Ahrefs study is a controlled experiment measuring causation. That is a harder standard than correlation, and it produced a null result.
Study 2: A 2025 arXiv preprint finds structured data is the third-strongest predictor of AI citation likelihood
A September 2025 preprint by Arlen Kumar and Leanid Palkhouski (arXiv:2509.10762) took a different approach. Instead of measuring what happened when schema was added, they audited 1,100 URLs that had already been cited across 1,702 citations from Brave Search, Google AI Overviews, and Perplexity.
Their regression found structured data as the third-strongest predictor of AI citation likelihood, associated with a +39% lift. The two factors ranked above it were metadata/freshness (+47%) and semantic HTML (+42%).
This is a cross-sectional correlation study, not a controlled experiment. It tells you that pages which get cited tend to have structured data more often than pages that do not. It does not tell you that adding structured data to an uncited page will cause it to be cited. That distinction matters.
The study has not yet been peer-reviewed (it is a preprint). Its platform scope also differs from the Ahrefs study: it includes Brave Search, which Ahrefs did not test.
Why do these findings diverge?
Two structural differences explain most of the gap.
1. Correlation versus causation. The arXiv preprint measures association: cited pages tend to have structured data. The Ahrefs experiment measures causation: adding structured data did not significantly change citation rates. Both findings can be true at the same time. Pages that were already well-structured (and thus citation-ready) may have had structured data as one of many co-occurring quality signals. Adding schema to a page that lacked it does not necessarily replicate all those other signals.
2. Platform and sample scope. Ahrefs followed pages that changed over time across Google AI Overviews, AI Mode, and ChatGPT. Kumar and Palkhouski analyzed citation patterns across Brave, Google AI Overviews, and Perplexity. Different platforms, samples, and study designs answer different questions, so their effect sizes are not directly comparable.
What the evidence does and does not support
The studies above do not prove schema markup has no effect on AI citations. They also do not prove it has a meaningful positive effect. Here is a fair summary of what the evidence currently justifies saying:
Supported:
- Schema markup is positively correlated with being cited by AI engines in cross-sectional data (Kumar & Palkhouski, 2025).
- Adding schema to pages already being cited does not appear to increase their citation rate (Ahrefs, 2026).
- Schema is a well-documented signal for traditional search features (rich snippets, knowledge panels, featured snippets) independent of the AI citation question.
Not supported:
- Schema markup predictably increases AI citation rates (no controlled study has found this).
- Schema is the primary driver of AI citation likelihood (even the arXiv preprint ranked two other factors above it).
How practitioners and tools approach this today
The schema debate has pushed practitioners toward a “low cost, uncertain upside” framing: add schema because it does not hurt and because it has confirmed benefits for traditional search, but do not treat it as a lever you can pull to move AI citation numbers.
Tools built around AEO take different positions here. HubSpot”s free AEO Grader surfaces schema as one of several structural signals without claiming a fixed citation multiplier. Qwairy Site Readiness surfaces site issues related to AI search, but it does not replace a full technical audit or implement schema. Surfer builds schema generation into its content workflow as a formatting baseline, alongside heading hierarchy and semantic structure.
Temso approaches it similarly: schema is part of the fix set that the platform generates alongside direct-answer formatting and FAQ structure, framed as content-ready infrastructure rather than a standalone citation lever.
Otto SEO automates schema implementation as part of technical SEO workflows, where the benefit case rests on traditional search features rather than AI citation uplift specifically.
The more productive framing across these tools is that schema is table stakes for technical correctness, while the higher-leverage moves for AI citation are content-level: direct answers in the first 40-60 words of a page, sequential heading structure, semantic HTML, and content freshness. The Kumar & Palkhouski preprint itself ranked metadata/freshness and semantic HTML above structured data in citation prediction.
If you want to track whether schema (or any other change) is actually moving your AI citation rate, you need a monitoring tool that samples the same prompts across multiple runs before and after any change. A single run is noise. See how to measure AI citation rates properly.
The bottom line
Schema markup does not appear to be the AI citation lever that much of the SEO industry assumed it would be. The Ahrefs controlled experiment is the strongest evidence for that position. The arXiv preprint provides the strongest evidence for the other side, though it measures correlation, not causation.
For your own program: add schema because it is correct technical practice and because it supports rich snippet eligibility. Monitor your AI citation rate across multiple engine runs before and after any schema change. And treat any headline number about schema and AI citations, including the ones on this page, as a lead for further investigation rather than a settled fact.
The full AEO tool comparison, including how each platform handles structured data in its workflow, is at /rankings/aeo-tools.
Methodology note
This page synthesizes a controlled experiment from Ahrefs and a cross-sectional arXiv preprint from Kumar and Palkhouski. No original data was collected for this piece. Each finding is attributed to its primary source. The preprint caveat is stated inline. For our full scoring framework, see /methodology.