Does Schema Markup Help AI Citations? What the Data Shows

Schema markup does not meaningfully move AI citations. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 controls, and citations barely budged: −4.6% on Google AI Overviews, +2.4% on AI Mode, +2.2% on ChatGPT. Only the negative was statistically significant. Adding structured data is not the citation lever the checklists promise.

What the Ahrefs schema study found

The study, published May 11, 2026 by Louise Linehan, is the cleanest test of the question so far. It isolated pages that had no JSON-LD, watched for the moment they added it, and compared their citation trajectory to matched control pages that never added schema. That before-and-after design is what separates it from the correlation studies everyone was passing around.

The result: adding schema produced no major uplift on any platform.

The only statistically real effect pointed the wrong way, and it was tiny. The two positives were noise. If schema were a citation lever, a controlled test of nearly 2,000 pages adding it would show something. It showed nothing worth acting on.

Then why do cited pages have more schema?

Because correlation is not causation, and the raw numbers are seductive. In the same research, Ahrefs looked at 6 million URLs and found that 53% of AI-cited pages carried JSON-LD, close to three times the rate of non-cited pages. Read fast, that sounds like proof schema helps. It is the opposite of proof.

Schema markup lives on better-maintained, more technically sophisticated sites. Those sites also publish stronger content, build more authority, earn more links, and rank well in classic Search. All of that drives citations. The schema is riding along as a marker of a healthy site, not doing the work. Strip out the confounders and the causal signal disappears, which is exactly what the controlled test showed. This is the same correlation trap that shows up in brand-mention data: a signal can predict citations without causing them.

The detail that explains the whole result

The study found AI systems extracted a page's visible HTML and ignored JSON-LD, Microdata, and RDFa when deciding what to cite. That single finding accounts for the flat numbers. Structured data is written for machines in the page's code, invisible to a human reader. The models assembling AI answers were reading the words on the page, the same text a person sees. Markup a human never reads is not what a generative engine quotes back.

That is not to say Google's systems never touch structured data. Google's rich-results pipeline uses it heavily. But the retrieval-and-generation step that picks AI citations is working off the rendered content, so hiding your key facts only inside JSON-LD does not get them cited.

Google says the same thing in its own docs

This is not a rogue vendor finding. Google's AI features documentation states it plainly: "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." No special schema, no llms.txt, no AI-only markup. The same guidance runs through Google's broader GEO advice: optimizing for AI search is optimizing for Search, full stop.

Independent measurement agrees on direction. Seer Interactive's 2026 analysis of AI Overview citations found FAQ schema and author-bio markup correlated slightly negatively with earning the top citation slot, another sign schema is hygiene rather than a growth tactic.

Schema is still worth keeping (for other reasons)

None of this means you rip out your markup. Schema still earns things AI citations do not replace:

• Rich results. Product cards, review stars, video previews, recipe cards, and organization panels all depend on structured data. AI Overviews can pull from those rich surfaces, so the markup helps indirectly.
• Entity clarity. Clean Organization, Person, and Product schema helps Google resolve who and what you are, which feeds the knowledge graph and, downstream, retrieval.
• Search hygiene. It is cheap, standard, and expected. Keep your schema matching the visible text on the page, as Google recommends.

Treat schema as table stakes for regular Search, not as a way to buy AI citations. The mistake is budgeting it as a citation tactic and expecting a lift that the data says will not come.

What to do instead

If citations are the goal, spend the effort where the evidence points. Cyrus Shepard's May 2026 meta-analysis, the most rigorous public attempt to weight AI citation factors by strength of evidence, puts the real levers at the top: URL accessibility, search rank, fan-out rank, preview control, and query-answer match. Schema is not among them.

1. Fix eligibility first. Make the page crawlable, indexable, and snippet-eligible. A nosnippet directive blocks AI citation faster than any schema fixes it.
2. Rank for the query and its fan-out. Top-10 placement for the seed term and the sub-queries engines fan out behind it is the strongest content-adjacent signal.
3. Answer the exact question near the top. AEO 101: put a clean, self-contained answer in the first breath so a model can lift it.
4. Earn brand mentions. Off-page, mentions across trusted sources track AI visibility far better than markup does.

Structured data belongs on your page. It is just doing a different job than the one the "add schema to get cited" advice sold you. For the tools that actually track whether engines cite you, see our AEO tools roundup, and for how the retrieval step picks sources in the first place, how ChatGPT picks its sources.