Does Schema Markup Help AI Citations? What a Controlled Study Really Found

Updated: July 2026. Somewhere in the last two years, “add schema markup” became the reflexive answer to every AI-visibility question. Not ranking in AI Overviews? Add schema. Not cited by ChatGPT? Add schema. The pitch is seductive because it’s concrete — you can ship a JSON-LD block this afternoon — and because a wave of secondary GEO blogs attached hard-looking numbers to it: “2.5× more likely to appear in AI answers,” “+40% AI Overview appearances,” “65% of ChatGPT-cited pages have structured data.” The problem is that when someone finally ran a controlled experiment, none of it held up.

This is the contrarian pillar in our schema cluster, and it answers a different question than the rest of it. If you want to know how to implement structured data — the eight JSON-LD types, the @graph pattern, the copy-paste code — start with our schema markup for AI visibility guide (implement it here) or the deeper schema for AI engines playbook. This article answers the prior question those guides assume: does schema markup actually help AI citations? The evidence says not directly — and the honest reframe is more useful than the myth.

The Myth: Schema as a Citation Turbo

The claim that structured data multiplies your odds of being cited by AI is everywhere, and it almost always arrives without a traceable source. Three figures in particular have circulated so widely they’re treated as settled fact: that schema makes you “2.5× more likely to appear in AI answers,” that “Tier-1 schema” yields “+40% AI Overview appearances,” and that “65% of pages cited by ChatGPT have structured data.” We went looking for the primary studies behind each. There aren’t any — every trail dead-ends at another blog quoting another blog.

Two of those numbers are worse than unsourced: they’re contradicted by the one controlled experiment that exists. And the 65% figure, even if it were real, would be a textbook correlation trap — the pages ChatGPT cites tend to be established, well-resourced sites, and established sites also tend to run schema. That co-occurrence tells you nothing about whether the schema caused the citation. This is the recurring failure mode of GEO advice: a plausible-sounding multiplier, no experiment underneath it, and a causal story bolted onto a correlation.

Why does the myth persist? Partly because schema does do real work elsewhere — it earns rich results, it helps Google understand your content — so “schema helps” feels intuitively true and gets over-extended to citations. And partly because it’s an actionable, satisfying answer to an anxious question. The corrective isn’t to declare schema useless. It’s to measure what it actually moves.

What the Ahrefs Study Actually Found

In 2026, Ahrefs ran the experiment the myth needed and never had. They identified 1,885 pages that added JSON-LD structured data between August 2025 and March 2026, then matched each one to comparable control pages — roughly 4,000 in total, drawn from the same domains, at the same pre-period citation level, that never added schema. Using a difference-in-differences design over a 30-day before/after window, they measured how AI citations changed for the schema-adopters relative to the controls. If schema were a citation lever, the adopters should have pulled ahead. They didn’t.

Read the numbers precisely, because the wrong summary writes itself. Google AI Mode came in at +2.4% and ChatGPT at +2.2% — both statistically indistinguishable from zero. In Ahrefs’ own words, there was “no citation growth in AI Mode, no citation growth in ChatGPT.” The only statistically significant movement was on Google AI Overviews, at −4.6% — a small decline, not a gain. Whatever else the data shows, it does not show schema lifting citations on a single engine.

There is a limitation you have to hold in the same hand as the result, and Ahrefs states it openly: every page in the dataset already had 100+ Google AI Overview citations in February 2025, before any schema was added. These were pages already deep in the AI consideration set. So the study answers a specific question — “does adding schema push already-visible pages higher?” — and the answer is no. It cannot tell you whether schema helps a page that AI systems have never discovered, and it should never be read as “schema hurts.” That −4.6% is a small drift on saturated pages, not evidence of harm.

Why Schema Doesn’t “Boost” Citations

The null result isn’t mysterious once you look at how engines read a page at citation time. A searchVIU experiment tested exactly this: it had five AI systems — ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode — fetch a page in real time and recorded what each one actually parsed. The finding is blunt. None of them extracted the JSON-LD. They read the visible, rendered HTML — the text a human sees — and ignored the structured data, hidden Microdata, and hidden RDFa entirely. Only Gemini executed JavaScript at all (surfacing 4 of 8 prices versus ChatGPT’s 3 of 8). At live-retrieval time, your schema block is not in the room.

That mechanism explains the Ahrefs numbers cleanly: if the model reads your visible copy and skips your JSON-LD when it fetches a page to answer a query, then adding JSON-LD can’t change what it cites. This is also why the “schema turbo” framing was always mechanically implausible — it assumed a consumption path the engines don’t use at that stage.

One honest caveat keeps this from over-claiming. “Engines ignore schema” is true at the live-fetch stage, but not necessarily everywhere. Some surfaces — Google AI Overviews and Bing Copilot among them — draw on a pre-built search index rather than doing a live fetch, and structured data is read when that index is built. So schema isn’t invisible across the whole pipeline; it just isn’t a live-fetch citation lever. That nuance is the difference between “schema is useless” (false) and “schema doesn’t buy citations” (true).

What Schema Really Does: Entity Disambiguation

Here’s where the story stops being a debunk and becomes a repositioning — and the most credible voice for it is Google itself. Gary Illyes has been consistent for nearly a decade: back at Pubcon 2017 he told SEOs to add structured data because “during indexing, we will be able to better understand what your site is about… indirectly it leads to better ranks because we can rank easier.” He reaffirmed it at Search Central APAC 2025: schema is not a direct ranking factor, but it is essential for entity understanding — resolving which real-world entity your content refers to. His own example: distinguishing Rome the city from Roma the football club.

John Mueller draws the same line from the other side. He has repeatedly stated structured data is not a ranking factor, and Google’s own policy proves where schema sits: misuse it and you lose your rich results, not your rankings. There is no scoring penalty because schema was never in the scoring layer to begin with. It lives in the understanding-and-eligibility layer — which, for AI search, is exactly the layer that matters.

The mechanism is concrete markup you already know. sameAs links your entity to authoritative external profiles — your Wikipedia or Wikidata entry, your official social accounts — so a system can confirm “this page and that profile describe the same real-world thing.” @id gives that entity a stable identifier engines can consolidate across pages and sources. Together they’re the documented way to separate your brand or person from same-named others and to reinforce trust. This is the layer where an AI engine resolves who you are — and an entity a model can’t identify is one it has no basis to trust or cite with confidence.

That’s the honest bridge back from the Ahrefs limitation. The study couldn’t test schema’s role in first-time discovery or disambiguation because its pages were already known entities. But entity resolution is precisely where schema earns its keep in AI search — the foundation, not the accelerant. It’s also exactly the job of Rankeo’s Entity Registry: maintaining a single, stable @id per entity and stitching it — via Schema-Stitch — to a coherent knowledge graph so engines see one verifiable identity instead of a scatter of same-named fragments.

When Schema Still Matters (a Lot)

None of this is a case for ripping out your structured data. Schema remains a legitimate technical foundation with measurable payoffs — they just aren’t AI-citation payoffs, and keeping the metrics separate is what keeps you honest.

The clearest evidence comes from classic search. Schema App documented what disambiguation does when you measure it properly: on a set of location pages (11 test versus 4 control, over 85 days), adding spatialCoverage and sameAs to disambiguate the locations drove +46% impressions on location queries and +42% clicks on non-branded queries. A healthcare blog that added entity links via mentions and sameAs saw +86.75% total queries and CTR lifts above 300% on target terms. Read those numbers for what they are: Google Search impressions and clicks — not AI citations. They show entity linking paying off through disambiguation and relevance, which is the point. They are not evidence that schema moves AI citations, and importing them into that claim would smuggle the myth back in.

So keep schema for the jobs it genuinely does: eligibility for rich results in classic search, entity understanding that helps engines index and consolidate you, and the verifiable identity that AI systems need to trust a source. Ship it well — a clean @graph with stable @id references and honest sameAs links — and it’s a foundation. Ship it expecting a citation spike and you’ll be disappointed by the data.

What to Do Instead to Get Cited

If schema is the verifiability foundation rather than the citation lever, the work splits cleanly in two — and you need both halves.

Earn verifiable identity. This is the half schema supports. Give every important entity a stable @id, link it with accurate sameAs references to your authoritative profiles, and name yourself consistently everywhere so an engine consolidates the fragments into one recognizable source. This is entity work, not markup-for-markup’s-sake — the discipline our entity SEO guide covers end to end, and the reason a model can trust that the page it’s reading belongs to who it claims to.

Write content a model will actually cite. This is the half schema does nothing for, and it’s where citations are truly won. Because engines read your visible text at retrieval time, the leverage is in the prose: front-load the specific answer, use definitive language, make each section self-contained enough to lift without stitching five tabs together, and build genuine topical depth. This is the practice of GEO — optimizing for generative engines by earning citations on merit, then probing the engines with real queries to see whether you actually show up.

The two halves compound. A verifiable entity that publishes weak content won’t get cited; strong content from an entity the model can’t identify gets attributed to someone else or not at all. Schema is necessary plumbing for the first half and irrelevant to the second — which is why “just add schema” was never going to be the answer to “why aren’t we cited?”

The Verdict

Does schema markup help AI citations? Not directly — and the honest version of that answer is more useful than the myth it replaces. A controlled Ahrefs study found no citation uplift from adding JSON-LD (AI Mode and ChatGPT in the noise, AI Overviews a small −4.6% on pages already saturated with citations), the mechanism explains why (engines read your visible HTML at retrieval, not your schema), and Google’s own line — from Illyes and Mueller — confirms schema was always an entity-understanding tool, not a scoring lever. So drop the “schema turbo” framing and keep the real value: schema earns you a verifiable identity via @id and sameAs, the layer where an AI engine decides whether it can trust who you are. That’s a foundation worth building well — and it’s exactly the layer Rankeo’s Entity Registry operationalizes.

FAQ