AEO

Schema Markup and AI Citation: What a Test Found

Schema markup AI citation advice has one controlled test behind it. Ahrefs tracked 1,885 pages and found no lift. See what schema is for.

By David Jubé · Aug 7, 2026 · 13 min read

TDM Insights report showing 1,885 pages and no citation lift.

Table of Contents

Schema markup AI citation advice has exactly one controlled test behind it, and that test measured no lift on any engine it tracked.

Does adding JSON-LD to a page make an AI engine cite it more often?

Ahrefs matched 1,885 pages that added schema against 4,000 control pages that did not, then counted citations for 30 days on either side of the change. No lift on Google AI Mode, none on ChatGPT, and a 4.6 percent decline on Google AI Overviews.

None of that makes the markup useless. It does mean the advice to add schema for AI visibility was built on a correlation, and the correlation survived the test while the advice did not.

Key takeaways

Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, each matched against control pages that added nothing.
Google AI Mode came in at 2.4 percent and ChatGPT at 2.2 percent, both close enough to zero to be noise.
Google AI Overviews fell 4.6 percent against its controls, a real gap that nobody has explained.
Every page in the sample already had more than 100 AI Overview citations before schema was added, so the test says nothing about pages AI has never seen.
Schema keeps its older job of rich results and machine-readable identity, and that job is narrowing as Google retires markup types.

Adding Schema to 1,885 Pages Moved Nothing

No, adding schema markup does not get a page cited more often by AI, on the only test that isolated the question. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, gave each one three control pages from other domains with similar prior citation levels, and compared the 30 days before the change against the 30 days after.

That is the direct answer.

The control group is what makes it worth anything. Citations were moving hard across AI search in that window, with AI Overviews contracting while AI Mode expanded, so a plain before-and-after would have measured the platform rather than the markup.

Raw AI Mode growth on the treated pages came in at 43 percent, which reads as a result until the control pages show almost the same gain.
Strip the platform trend out and AI Mode lands at 2.4 percent, inside the margin of error.
Google AI Overviews moves the other way, a 4.6 percent decline, and it is the only figure in the study that clears statistical significance.

The result is flat. What is worth another minute is why so many people expected it not to be.

Schema Markup AI Citation Advice Started as a Correlation

The same write-up opens with the finding that started all of this: across 6 million URLs, pages cited by AI were almost three times more likely to carry JSON-LD, and 53 percent of AI-cited pages had schema on them.

That gap is real, and it is the whole evidentiary base for most of the advice in circulation.

Correlation of that shape has an obvious rival explanation. Sites that bother with structured data also maintain their pages, publish more authoritative work, earn more links, and rank better in classic search. Markup arrives with a crowd, and the crowd does most of the lifting.

Expected: markup gives an engine a cleaner read of the page, so the marked-up page wins the citation.
Measured: the role structured data plays in AI visibility has been argued far more often than it has been tested, and the one test that isolated it found nothing.
Left standing: the sites doing schema well were already doing everything else well, and removing the markup leaves the rest of those signals intact.

A 53 percent figure travels well because it sounds like a threshold. It describes who already had markup, not who gets picked, and the distance between those two readings is the entire argument.

JSON-LD Is a Format, Not a Signal

JSON-LD is a block of machine-readable labels that sits alongside the visible page rather than inside it. Ahrefs tested that block and nothing else, which is the right thing to test and a narrower thing than the advice claims.

Format and effect get conflated constantly. Ahrefs’ own explainer walks through what schema markup is and how to add it, and adding it correctly has never been the disputed part. The disputed part is what a correct addition buys.

Nothing in the format itself promises a citation.

This is the first of five widely repeated claims the pillar puts through the same Verdict Model: claim, test, verdict, limit. Test is the step the advice skips, and skipping it turns a description of the sample into a prescription for everyone else.

Four Tests, Three Engines, One Small Decline

Ahrefs ran the data four ways rather than once: a two-sample t-test, a matched difference-in-differences analysis, a week-by-week event study, and a re-run of the difference-in-differences on a symmetrical window that excluded the recrawl period.

All four pointed the same way.

The three figures below come from the difference-in-differences test Ahrefs treats as its most reliable read, and Search Engine Roundtable reported the same topline independently without softening it.

Engine	Effect vs matched controls	How to read it	Source
Google AI Overviews	4.6 percent decline	Statistically significant, roughly 12 fewer citations a day on pages getting hundreds	Ahrefs matched difference-in-differences, May 2026
Google AI Mode	2.4 percent gain	Indistinguishable from zero once the platform trend is removed	Ahrefs matched difference-in-differences, May 2026
ChatGPT	2.2 percent gain	Indistinguishable from zero	Ahrefs matched difference-in-differences, May 2026

Odds of an AI Overviews gap that size appearing by chance run at about one in 2,500.
Both groups were already declining on AI Overviews before schema was added; the treated pages fell a little faster.
Raw AI Mode growth before controls was 43 percent, which is the number a study without a control group would have published.

Small, significant and unexplained is an uncomfortable combination to report. It is also the honest description, and Ahrefs states it that way rather than reaching for a mechanism.

What This Test Cannot Tell You

It cannot tell you schema has no effect on AI citation. It can tell you that on these pages, in this window, with these markup types pooled together, no effect showed up.

Every page in the sample already had more than 100 AI Overview citations before any markup was added.

So the test measured whether schema lifts a page engines were already citing. It did not measure whether schema helps a page get noticed for the first time, which is the position most founders are writing from.

Pooled types: Article, FAQ, Product, HowTo and Organization were treated as one category, so a real effect from one type could be cancelled by drag from another.
Thirty days: an effect that builds over 60 or 90 days would not appear in the window measured.
HTML only: markup injected by JavaScript sat outside the test, and crawlers treat the two differently.
Co-occurring changes: pages that add JSON-LD often ship content and template changes in the same deploy.

The Amplifier Argument, Stated Fairly

There is a serious counter-position here, and it deserves the same treatment as the claim it opposes. Schema on this reading is not a trigger but an amplifier: on a page carrying genuine ambiguity, it removes enough of that ambiguity to tip a borderline evaluation.

Ahrefs tested pages well past the borderline stage.

If the amplifier argument holds, a test built on already-cited pages could not have caught it. Absence of a measured lift is not proof of no effect, and treating it as one would be the same overreach pointed in the opposite direction.

The defensible position is narrower than either camp wants. Schema does not lift pages that engines already cite. What it does for a page engines have never seen is open, unresolved, and worth someone’s next study rather than a confident line in a pitch deck.

What the null result does is move the question. If markup is not what an engine reads at the moment it picks a source, something else is.

Retrieval Reads What a Reader Reads

During direct retrieval, five AI systems pulled visible HTML and ignored the markup entirely. Search Engine Journal summarised the searchVIU experiment alongside the Ahrefs data: ChatGPT, Claude, Perplexity, Gemini and Google AI Mode all extracted rendered content, while JSON-LD, hidden Microdata and hidden RDFa went unused.

That covers one stage of the pipeline, not the whole of it.

Indexing and entity understanding happen earlier, and markup could still be doing work there. What the experiment rules out is the picture most of the advice assumes, in which an engine reads your JSON-LD at answer time and decides you are the better source.

Loose Citation Is Not a Careful Read of Your Markup

Citation itself is looser than the advice imagines. An academic audit published in 2023 tested four generative search engines and found that only 51.5 percent of generated sentences were fully supported by their citations. Citation precision ran at 74.5 percent, so a quarter of citations did not support the sentence they were attached to.

Systems citing that loosely are not running a precise read of anyone’s markup at answer time.
The engines in that audit have changed since 2023, so treat the figures as a snapshot rather than a constant.
The stage order is the durable part: content is retrieved, an answer is generated, sources are attached last.

Writing the answer into the visible page is the part you control, and it is where I would put the hour. I have covered the mechanics separately in a guide on structuring a page so it can be quoted and one on the formats LLMs pull from most.

Founder Insight Schema is a labelling job, not a visibility job, and the quickest way to price it correctly is to ask what breaks if you remove it. If the answer is a rich result or an entity link rather than a citation, you have your budget.

Rich Results Are the Case for Schema, and That Case Is Narrowing

Schema still belongs on the page for reasons that never depended on AI citation. Google’s structured data guidelines are built around helping Search understand a page and, where a site qualifies, show a richer result.

That case is real, and it is smaller than it was two years ago.

Google stopped showing the FAQ rich result on 7 May 2026 and removed the documentation for it the following month. Search Engine Roundtable logged the drop as it happened, and HowTo, Course Info, Claim Review and Estimated Salary went the same way, each retired once it became a familiar tactic.

The pattern is worth naming because it repeats. A markup type ships with a visible reward, the reward gets chased at scale, Google withdraws it, and the type stays valid while the reason people adopted it disappears.

Any markup tied to a single search feature should be treated as temporary. Anything that labels what a page is, rather than how it looks in a result, tends to survive.

What remains is worth doing on its own terms.

Article markup: labels headline, author and publish date against the schema.org Article type, and that job never depended on an AI engine reading it.
Organization markup: ties every page to one machine-readable identity, per the schema.org Organization type, which is entity work rather than citation work.
FAQ markup: stays valid and still gets parsed, because Google says it will keep using the data to understand pages even with the rich result gone.

Where the Markup Still Earns Its Hour

Rich-result eligibility and entity clarity are two different jobs, and only one of them has been shrinking. I set out the classic search case in a separate piece on structured data as a founder’s 80/20, and the implementation detail sits in the answer-first writing guide.

None of that is a workaround for the test result. It is a separate case, made on separate grounds. It survives because it was never the claim the test checked.

My Posture: Ship the Markup, Stop Selling It as Visibility

Schema goes on every site I build, and I no longer describe it as a way to get cited. That is the verdict: the citation claim fails its only test, and the markup keeps its older, smaller job.

Ahrefs’ matched test is the reason for that change, not a preference. Markup that moved nothing across 1,885 pages has no claim on a visibility pitch.

Coverage is also not the same measurement as citation. A CMS report saying markup is complete tells you what went into the template, not what an engine did with it.

Audit what the CMS says is finished before you cite it, including your own past work.
Track citations rather than markup coverage, since Bing Webmaster Tools now reports which URLs get referenced in AI answers.
Spend the schema hour once, then spend the next ten on entity clarity and on content an engine can quote.

llms.txt runs the same shape: adoption first, confident advice second, no test underneath either. Until a test exists, schema stays on the page and stays off the pitch.

Frequently Asked Questions

Does schema markup help a page get cited by AI?

Not on the evidence available. Ahrefs matched 1,885 pages that added JSON-LD against control pages that added nothing and found no meaningful citation lift on Google AI Overviews, Google AI Mode or ChatGPT. That is one test on already-cited pages, so it rules out a large effect rather than every effect.

What did the Ahrefs schema study measure?

Citation counts before and after 1,885 pages added JSON-LD between August 2025 and March 2026, each treated page matched against three control pages from other domains with similar prior citation levels. Ahrefs measured 30 days either side of the change across Google AI Overviews, Google AI Mode and ChatGPT.

Why did AI Overview citations dip after pages added schema?

Nobody knows, including Ahrefs. The 4.6 percent decline is statistically significant but small, roughly twelve fewer citations a day on pages already getting hundreds, and both treated and control pages were falling before the markup went on. A content or template change shipped in the same deploy could account for it.

What is schema markup still worth doing for?

Rich-result eligibility where a site qualifies, and giving search engines a clean machine-readable read on a page’s identity. Article markup labels headline, author and date. Organization markup ties pages to one entity. Neither job depends on whether an AI engine cites the page more often.

Which schema types are worth keeping on a site?

Article and Organization carry the most durable value, since they label authorship and identity rather than chasing a visual feature. FAQ markup stays valid and parsed even though Google stopped showing the FAQ rich result in May 2026. Treat any type tied to a single SERP feature as temporary.

How do I check whether AI engines cite my pages?

Use a tool that reports citations directly rather than inferring them from markup coverage. Bing Webmaster Tools reports which of your URLs get referenced in Copilot and Bing AI answers. Third-party visibility trackers cover ChatGPT and Perplexity. Check citations, not the presence of JSON-LD.

Continue Reading

More From This Series

More from TDM Insights

Explore TDM Insights Topics