Does Schema Markup Increase AI Citations? What a Controlled Study Found

Quick answer: For pages that AI already cites, adding JSON-LD schema produced no meaningful citation lift. A matched difference-in-differences study tracked 1,885 pages that added JSON-LD against 4,000 control pages and measured Google AI Overviews at −4.6%, Google AI Mode at +2.4%, and ChatGPT at +2.2% — the last two statistically indistinguishable from zero [1]. Google's own documentation states there is no special schema.org markup you need to add to appear in AI features [2]. Schema still earns its place for rich results and entity recognition. Treat it as hygiene, put incremental GEO budget into crawlability, visible text, and passage-level clarity, and verify on your own pages with a controlled test.

If your GEO plan for this quarter is "ship more schema," this article gives you the evidence to re-price that bet.

---

What did the study actually measure?

Ahrefs published the study on May 11, 2026. It began with a correlation that has circulated widely in conference decks: across an analysis of 6 million URLs, pages cited by AI were almost three times more likely to carry JSON-LD than pages that were never cited. Among AI-cited pages, 53% were running schema [1].

That gap invites an obvious conclusion. The study then tested whether the conclusion survives a controlled design.

The team identified pages where schema appeared for the first time. Using HTML history from their crawler database, they labeled whether each URL contained a JSON-LD script tag and found the date it flipped from absent to present. That produced 1,885 pages that introduced JSON-LD between August 2025 and March 2026. Each treated page was matched to 3 control URLs from different domains with similar pre-period citation levels that never added schema, for roughly 4,000 controls. Citations were measured across the 30 days before and 30 days after the treatment date [1].

The matching step is what makes the result readable. AI citation volumes were moving sharply during this window on their own: AI Overviews were contracting while AI Mode was expanding. A simple before-and-after comparison would have measured the platform trend. The raw before-and-after growth for AI Mode came in at +43%, and control pages gained almost as much — which is how a headline number shrinks to +2.4% once the platform trend is stripped out [1].

The −4.6% deserves a careful reading. In absolute terms it works out to an average loss of about 12 daily citations per page, in a sample where most pages were receiving hundreds. Both treated and control pages were already on a steep downward path before schema was added. The authors state plainly that they cannot determine from this data whether schema caused the gap or whether other factors did [1].

Four separate analyses were run: a two-sample t-test, the difference-in-differences test the authors trust most, an event study checking whether the two groups were already drifting apart beforehand, and a repeat of the difference-in-differences using a symmetrical window that excluded the recrawling period. All four pointed the same direction — no citation growth on AI Mode, none on ChatGPT, and a real but small AI Overviews decline that cannot be pinned on schema with confidence [1].

---

Why did the 3x correlation look so convincing?

The correlation is genuine. The interpretation is where it breaks down.

Schema markup tends to live on better-maintained, more technically sophisticated sites. Those same sites publish stronger content, build more authority, earn more links, and maintain their pages. AI systems are more likely to retrieve exactly that kind of content, so cited pages over-index on many favorable signals simultaneously. Strip schema out and the rest of those signals very likely still carry the page through to a citation [1].

An independent cross-platform study reached a compatible conclusion through a different route. Its within-Google diagnostic reported that schema prevalence among AI-cited and non-cited Google pages was statistically indistinguishable — 43.1% versus 44.8% — which collapses the apparent effect [3].

Two studies, two methods, one direction. That is the standard of evidence worth acting on.

---

Do AI systems read schema at all?

This is where the picture gets more precise, and where an overcorrection would be a mistake.

A German SEO firm ran eight controlled tests in October 2025 across five AI systems, using a purpose-built page for a fictional product with prices deliberately placed in different locations: visible HTML, JavaScript-rendered content, JSON-LD only, JSON-LD injected via JavaScript, hidden Microdata, visible Microdata, hidden RDFa, and visible RDFa [4].

The core finding: the price that existed only inside JSON-LD was found by none of the five systems. Hidden Microdata and hidden RDFa were also ignored across the board. Where Microdata and RDFa were visible in the HTML, ChatGPT and Gemini did find the values — because the content was visible, and the markup format was incidental [4].

The authors are explicit about scope, and this qualifier matters. Their tests primarily observe the direct fetch phase, when a chatbot retrieves a page live. Structured data may well be extracted during training, indexing, and search phases — and index-based systems such as AI Overviews and Copilot have access to structured data stored in a search index [4]. No study reviewed here demonstrates that schema is inert everywhere in the pipeline. What the evidence supports is narrower and still actionable: content that exists only in markup is content most systems will miss at retrieval time.

The same tests produced a second useful result. Gemini executed JavaScript during live fetch and found the JS-rendered price, while ChatGPT, Claude, and Perplexity did not capture JS content live. Both index-based systems found the JS-rendered price after indexing, indicating their crawlers do execute JavaScript during indexing [4].

---

What does Google actually say?

Google's documentation on AI features answers this directly. Two statements are relevant, both from the page itself, last updated December 10, 2025:

There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.

You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add.

The same page lists the fundamentals that do apply: allow crawling in robots.txt and at the CDN or hosting layer, make content findable through internal links, provide good page experience, keep important content available in textual form, and make sure structured data matches the visible text on the page [2].

That last item is worth holding onto. Google's guidance positions schema as something that should agree with your visible content — which is a consistency requirement, and a different proposition from schema functioning as a citation lever.

---

So where should the budget go instead?

Here is the practical sequence. Start from what you already have, identify what is missing, and spend where the evidence is strongest.

What you probably already have. If your site is well maintained, you likely already have schema. Keep it. Rich results, voice assistants, knowledge graphs, and downstream entity recognition are all legitimate reasons to run JSON-LD [1], and Google expects it to match your visible text [2]. Removing working schema to chase this finding would be the wrong move.

What is more likely to be missing. Retrieval access and passage-level clarity. The evidence base is considerably stronger here.

One nuance holds the two findings together. A page-quality audit framework found that among 16 pillars, structured data ranked among those most strongly associated with citation [7]. That association is real and consistent with the 3x correlation. The controlled tests tell you what happens when schema is the only thing you change. Both can be true at once, and the combination is the actual lesson: schema travels with pages that get cited, and adding it alone does not move them.

---

How do you test this on your own site?

Aggregate findings describe averages across thousands of URLs. Your site is one case. A small controlled test is the only way to know what holds for you, and the study authors recommend exactly this [1].

1. Pick 5–10 test pages where you plan to add JSON-LD. Choose pages already receiving some AI citations so you have a baseline. Pages with zero citations make the result hard to interpret.
2. Pick 5–10 control pages with similar citation levels that you will leave alone. This step separates "schema did something" from "the platform shifted for everyone that month."
3. Record baseline citations for both groups across AI Overviews, AI Mode, and ChatGPT.
4. Add schema to the test pages only, note the date, and change nothing else during the window.
5. Compare both groups after 30 days or longer. The question is whether treated pages rose more than controls did.

Step 3 is where most teams stall, because standard analytics does not report citations by engine. Whatever setup you use, three capabilities decide whether a test like this is even readable:

Innflows is built for this layer. It audits whether AI crawlers can reach and parse your pages, checks technical readability, content structure, and entity consistency, and tracks brand citations engine by engine — covering ChatGPT, Google AI features, Perplexity, and Copilot alongside Chinese-language engines including DeepSeek, Doubao, Qwen, and Kimi. Once per-engine baselines exist, a controlled test like the one above becomes a routine quarterly measurement.

The audit half matters here as much as the tracking half. If the pages you are testing turn out to be blocked at the CDN, dependent on JavaScript for their key facts, or inconsistent between structured data and visible text, then schema was never the variable worth testing on them — and those are exactly the issues an access-and-readability audit surfaces before you spend a quarter measuring the wrong thing.

---

When does this finding not apply?

Three boundaries, stated plainly.

Pages AI has never cited. Every page in the study had 100+ AI Overview citations before schema was added. These pages were already inside the consideration set. For pages AI has never surfaced, schema may still help them get crawled, parsed, or indexed in the first place — the study cannot speak to that [1].

Specific schema types and longer windows. All schema types were pooled together: Article, FAQ, Product, HowTo, Organization. Some may perform differently. The measurement window was 30 days, so a slow-burn effect over 60 or 90 days remains untested. Only JSON-LD in the page HTML was studied, and AI crawlers appear to treat JavaScript-injected schema differently [1].

Non-citation benefits. Rich results eligibility, entity recognition, and knowledge graph population are separate goals with separate evidence. This finding is about AI citations on pages that already have them.

---

FAQ

Should I remove schema markup from my site?

Keep it. The study addresses whether adding JSON-LD lifts AI citations on pages already being cited, and it found no meaningful lift [1]. Schema continues to serve rich results, voice assistants, knowledge graphs, and entity recognition, and Google's guidance expects structured data to match your visible text [2]. Removing functioning markup would forfeit those benefits for no measured citation gain.

Does Google use schema for AI Overviews?

Google's documentation states there is no special schema.org structured data you need to add to appear in AI Overviews or AI Mode, and no additional technical requirements beyond being indexed and eligible to appear with a snippet [2]. Separately, index-based systems do have access to structured data stored in a search index, so schema being available to a system is a different question from schema changing citation outcomes [4].

If 53% of AI-cited pages run schema, how can schema have no effect?

Both facts hold together. Sites that implement structured data also tend to invest in technical SEO, publish authoritative content, build links, and maintain their pages. Cited pages over-index on all of those signals at once, so schema appears alongside citations without producing them [1]. An independent study found schema prevalence among cited and non-cited Google pages statistically indistinguishable at 43.1% versus 44.8% [3].

Will AI read product prices from my JSON-LD?

In controlled tests across five AI systems, a price placed only in JSON-LD was found by none of them during live retrieval, and hidden Microdata and hidden RDFa were also ignored [4]. The practical rule is to put facts you want quoted into visible page text, and let structured data mirror them.

What should I do instead to get cited more?

Prioritize retrieval access and passage-level clarity: confirm AI and search crawlers can reach your pages, serve key facts in visible text, and structure content so individual passages answer questions on their own. A controlled test across six engines reported a 17.3% citation improvement from structural changes alone with meaning held constant [5], and an analysis of 17 million citations found AI assistants favor fresher content [6].

How long should my own schema test run?

Run at least 30 days, and longer if your citation volumes are low. The published study measured a 30-day window and flagged that a slower effect over 60 or 90 days would go undetected [1]. Keep a matched control group throughout, since AI citation volumes shift for reasons unrelated to anything you changed.

---

Bottom line

A matched difference-in-differences test on 1,885 pages found that adding JSON-LD produced no meaningful AI citation lift on Google AI Mode or ChatGPT, and a small AI Overviews decline the authors decline to attribute to schema [1]. Google's documentation says the same thing from the other side: no special schema.org markup is required to appear in AI features [2].

The 3x correlation was real and the causal story attached to it was borrowed. That distinction is the whole lesson, and it generalizes past schema — most GEO advice in circulation rests on correlation that has never been isolated.

Keep your markup, stop treating it as a citation lever, and move the incremental budget to crawl access, visible text, and passage structure. Then confirm it on your own pages with a treated group, a control group, and per-engine baselines, because a finding that holds across thousands of URLs still has to hold on yours.

If you want to see where your own site currently stands on the two things this evidence does support — whether AI crawlers can reach and parse your pages, and which engines cite you today — start with an AI visibility audit and use its output to pick your test pages.

---

References

[1] - We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved. — Ahrefs, May 11, 2026

[2] - AI features and your website — Google Search Central, last updated December 10, 2025

[3] - Does Schema Markup Predict AI Citation? A Cross-Platform Empirical Study of Structured Data and Generative Engine Optimization — Kurt Fischman, SSRN

[4] - Schema Markup and AI: What ChatGPT, Claude, Perplexity & Gemini Really See — searchVIU, December 2, 2025

[5] - Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior — arXiv:2603.29979

[6] - AI Assistants Prefer to Cite "Fresher" Content (17 Million Citations Analyzed) — Ahrefs

[7] - AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO-16 Framework — arXiv:2509.10762