Adding schema markup did not increase AI citations in the best controlled test available. In a 2026 Ahrefs study, 1,885 pages that added JSON-LD structured data were matched against 4,000 comparable pages that didn't; citations to the schema pages moved -4.6% on Google AI Overviews (statistically significant), +2.4% on AI Mode and +2.2% on ChatGPT (neither significant). Yet across ~6 million URLs, AI-cited pages were about 3× more likely to carry schema. Both are true, and the gap between them is the whole story.
This is a dated reference, not a new dataset. Last reviewed: 6 August 2026. All figures come from Ahrefs' controlled schema experiment, published May 2026, which measured citation change over roughly a 30-day window per page across Google AI Overviews, Google AI Mode and ChatGPT. It is single-vendor and carries the caveats set out below, so read the direction as firmer than any one percentage and cite "Ahrefs, 2026 schema study" with the date when you reuse a figure. It confirms and sharpens, rather than contradicts, our earlier read that structured data is machine-legibility hygiene, not a citation lever, in prioritise your structured data and the three AI-citation levers ranked by evidence.
What did the schema study actually measure?
The causal question, not the usual correlation. Most "schema helps AI" claims rest on the observation that cited pages tend to have markup. Ahrefs instead used a before-and-after design: take pages that added JSON-LD in a known window, compare their citation change against matched pages that didn't, and see whether the markup itself moved anything.
| Attribute | Detail |
|---|---|
| Treated pages | 1,885 that added JSON-LD schema |
| Control pages | 4,000 matched pages that did not |
| Schema-add window | August 2025 – March 2026 |
| Engines measured | Google AI Overviews, Google AI Mode, ChatGPT |
| Observation window | ~30 days per page |
| Correlation sample | ~6 million URLs |
| Published | May 2026 (Ahrefs) |
Source: Ahrefs, 2026 schema study. The design matters because it isolates the effect of adding schema from everything else that makes a page citable, which a raw "cited pages have more schema" count cannot do.
Did adding schema move citations on any engine?
No engine showed a meaningful uplift, and one showed a small decline. Measured against the matched controls, the change in citations for pages that added schema was:
| AI engine | Citation change after adding schema | Statistically significant? |
|---|---|---|
| Google AI Overviews | -4.6% | Yes |
| Google AI Mode | +2.4% | No |
| ChatGPT | +2.2% | No |
Source: Ahrefs, 2026 schema study; changes are relative to 4,000 matched control pages. The two small positive moves sit inside the noise band, so the honest reading is "no detectable effect." The AI Overviews decline is real in the data but small, and most plausibly reflects other shifts over the window (Google reduced some rich-result surfaces in this period) rather than schema actively hurting you. The safe conclusion is the null one: adding markup did not buy citations.
So why are cited pages more likely to have schema?
Because schema travels with quality, not because it creates citations. This is the reconciliation the whole topic turns on, so state it plainly:
- The correlation is real. Across ~6M URLs, AI-cited pages were ~3× likelier to contain JSON-LD. A 2026 correlational read of on-page signals points the same way.
- The cause points the other way round. The sites that add schema are disproportionately well-resourced, well-structured, strongly-corroborated sites, exactly the ones that earn citations for reasons that have nothing to do with the markup.
- The controlled test breaks the tie. When you hold quality roughly constant by matching pages and change only the schema, the citations don't follow. That is the difference between "cited pages have schema" and "adding schema makes pages cited."
Schema is a marker of a well-built page, not a mechanism that earns the citation. Cited pages carry markup for the same reason they carry good content, the underlying quality, so adding markup to a page without that quality changes little.
Confusing the two is the single most common error in reading AI-visibility studies, and it is worth its own method: see how to tell whether an on-page change actually moved your AI citations and the plain-English definition of correlation vs causation.
What are the study's limits?
Real ones, which is why the claim is "not a lever," not "never matters." Read the numbers with these in view:
| Caveat | Why it matters |
|---|---|
| Only already-cited pages | Every treated page had 100+ AI Overview citations in Feb 2025, so the test asks "does schema add citations to pages that already win?", not "can it help a page break in?" |
| Schema types pooled | Product, FAQPage, Article and the rest were grouped, so a type that helps could be masked by types that don't |
| Equal-weighted citations | A citation on a high-value query counts the same as a trivial one |
| ~30-day window | A short observation period given the real crawl-to-cite lag |
| Single vendor, three engines | Ahrefs' measurement, US-weighted, not the whole AI-answer ecosystem |
Source: Ahrefs, 2026 schema study, with limitations noted by Ahrefs and industry commentary. None of these rescue schema as a growth lever; they bound the claim to "adding markup to a decent page is not a reliable citation gain," which is exactly the useful conclusion.
How does this square with Google, and with our own guidance?
Cleanly, and that is the point of writing it down. Three sources now say the same thing from different angles:
- Google's official guide states you don't need special files or markup to appear in its AI features, that it's still SEO. This controlled test is the empirical version of that statement for schema specifically.
- Our levers piece ranked structure and authority as the signals that move citations and flagged that the schema link was correlational, in freshness, structure, authority. The study confirms schema is not itself the structural lever.
- Our structured-data playbook already told you to add the right types and then measure whether it moved anything, in prioritise your structured data. The study is the large-sample answer to that measurement: mostly, it won't move citations on its own.
There is no contradiction to resolve, only a sharpening: schema is table-stakes machine-legibility that helps engines parse and reuse your facts, and it is worth keeping accurate, but it is not the thing that earns the citation.
What should you do with structured data now?
Keep it, right-size it, and stop treating it as a growth lever. Concretely:
- Keep accurate schema that matches your visible content. It labels your facts for machines, still drives non-AI rich results, and is cheap hygiene, the case in answer-engine optimization is unchanged.
- Don't expect uplift from adding more. On this evidence, bolting extra markup onto an already-decent page is not a reliable way to gain AI citations; budget accordingly.
- Spend the freed effort on the real levers. Non-commodity content, entity strength, and third-party corroboration are what move citations; schema just makes the facts on those pages legible.
- If you do add markup, test it honestly. Baseline, add, and read a smoothed trend against a control, the method in testing whether a change moved your citations, rather than trusting a single before-and-after.
The published averages tell you schema is not the lever; the number that governs your strategy is whether the changes you do make actually move how the engines cite you, measured across engines and over time. That per-change, per-engine citation measurement is exactly what Buffy Intel is built to provide. Questions: [email protected].