← Musings

Does Schema Markup Actually Get You Cited in AI Search?

6 July 2026

I read all ten studies that have measured whether schema markup wins AI citations, and checked every number against its original source. The studies that found an effect skipped the controls. The best-controlled found nothing, or a small negative.

AI searchSchemaResearch

There’s a statistic that travels through conference talks and LinkedIn posts: sites with structured data and FAQ blocks saw a 44% increase in AI search citations, attributed to BrightEdge. I went looking for the study. It doesn’t exist. BrightEdge has published a real 44% statistic — that Google’s AI Overviews are 44% more likely to criticize brands than ChatGPT — which has nothing to do with schema. Somewhere along the way, the real number was lifted from one claim and attached to another.

That search became a bigger one. I ran a systematic review of everything measurable on one question: does adding schema.org markup to a page make AI search products more likely to cite it? Five reproducible arXiv queries (71 unique papers screened), targeted searches, citation chasing through the reference lists, and a rule borrowed from the replication literature: no claim counts until the primary source is fetched and the figure confirmed. Ten sources measure the schema-to-citation relationship directly. Graded and plotted, they produce the cleanest pattern I’ve seen in this field.

less controlledmore controlledFischman 2026null / negative · GEE, rank + clusteringAhrefs DiDnull / negative · matched controlsOtterly canarynull / negative · schema-only fact testSearchAtlasnull / negative · coverage bucketsAirOps × Indigpositive · balance checks, no rank controlGEO-16positive · cited-only corpusDigital Appliedpositive · no controlsSchema Apppositive · own-site case studyEvertunepositive · no methodologyBrightEdge "44%"positive · no locatable study
Ten direct sources on schema markup vs AI citation, ordered by methodological control. Green = reported a positive schema effect; grey = null or negative. No positive result survives past the point where rank or design controls appear. Grades follow the review's A–D scheme.

Every study that found schema helps is missing the controls. Every study with controls found nothing — or a small negative.

The three positives first, because they deserve to be taken seriously before being taken apart. AirOps and Kevin Indig analyzed 16,851 queries against ChatGPT and found pages with JSON-LD cited at 38.5% versus 32.0% without — a 6.5-point gap, and they did real work checking that schema pages weren’t simply longer, better-matched, or higher-authority. What they didn’t control is retrieval rank. GEO-16, an arXiv preprint behind the “+39% structured data lift” you may have seen quoted, has a subtler problem: its 1,100-page corpus was harvested from citations, so every page in the study was already cited by at least one engine. There is no population of uncited pages anywhere in the analysis — the comparison that would make the number meaningful was never run. And Digital Applied’s 5,000-site audit reports a 0.34 correlation between valid schema and AI citations with no controls at all, though it contributes one genuinely useful field fact: 49% of deployed schema in the wild fails validation.

Now the controlled side. Ahrefs tracked 1,885 pages that added JSON-LD against ~4,000 matched controls: AI Overviews citations fell 4.6% relative to controls, and the ChatGPT and AI Mode movements were statistically zero. SearchAtlas bucketed domains by schema coverage and found visibility distributions “highly similar” across every bucket on three platforms. Otterly planted a fact that existed only inside FAQ schema on their own site; no AI platform ever used it. And the pivotal study — Kurt Fischman’s cross-platform analysis of 730 AI citations, which I verified against the full 24-page paper — is practically a lesson in why the positives look positive. His naive pooled result showed schema negatively associated with citation. After correcting for how the control set was built and clustering errors by query, the association collapsed to null. The strongest predictor left standing wasn’t schema at all.

Fischman 2026Google organic position → AI citation rate43%127%220%310%55%7AirOps × IndigChatGPT retrieval position → citation rate58.4%154.4%235.5%329.9%424.6%614.2%10
Citation rate by retrieval position, measured independently in two studies. Left: Fischman 2026 — Google organic rank vs citation by ChatGPT and Gemini (positions 1–7). Right: AirOps × Indig — ChatGPT's own retrieval order (positions 1–10). Different engines, different methods, the same collapse. Any schema study that ignores rank is mostly measuring rank.

That predictor was the page’s position in the search results — its rank. In Fischman’s data, a page at position 1 in Google’s organic results was cited in 43% of the queries where it appeared, falling to 5% by position 7. AirOps found the same gradient independently, on ChatGPT’s own retrieval: 58% citation at position 1, 14% at position 10. Two studies, different engines, different methods, same shape. AI citation is substantially a downstream echo of ordinary retrieval ranking — which means any schema study that doesn’t control for rank is mostly measuring rank. That is precisely the control the positive studies lack.

One finding in this research gets misquoted more than any other, so it is worth spelling out in full. Fischman found “attribute-rich” schema — Product and Review markup with real prices, ratings, and specs — cited at 61.7% versus 41.6% for generic schema like Article and BreadcrumbList. That 20-point gap is quoted everywhere as evidence that rich schema wins. Here’s what the full paper shows and the summaries omit: pages with no schema at all sat at 59.8%, statistically indistinguishable from attribute-rich. Generic schema performed worse than having no markup at all.

Attribute-rich schemaProduct/Review, populated values61.7%No schema at allstatistically the same (p = .71)59.8%Generic schemaArticle / Organization / Breadcrumb41.6%
Fischman 2026, citation rates by schema condition — the full comparison. The widely quoted result is the 61.7% vs 41.6% gap (p = .012). The omitted bar is no-schema at 59.8%: indistinguishable from attribute-rich (p = .71). Rich markup recovers the baseline; generic markup sits below it.

Read correctly, the field’s one live pro-schema signal says: rich markup gets you back to where having no markup already puts you, and boilerplate markup may carry a small penalty. And even that result has a simpler explanation: pages with populated price and rating fields also display those facts in visible text, and a peer-reviewed SIGIR 2026 experiment (252,000 controlled trials across six models) found that visible price and recency information act as citation gatekeepers with enormous effects, while formatting-only changes do nothing. The visible facts are what earn the citation. Whether the markup wrapping them contributes anything is exactly what no study has isolated.

“Schema does nothing” would still be too strong, because “AI visibility” isn’t one mechanism — it’s a pipeline, and schema could plausibly act at four different points in it.

stage 1
Crawl / index
Google & Bing parse structured data
plausible
stage 2
Retrieval
rank dominates; markup fields lift retrieval only in lab pipelines built to index them (SAGEO)
open
stage 3
Read / chunk
markup stripped by most pipelines
evidence: no
stage 4
Generate / cite
models cite what reaches context
no direct path
Where schema could act in an AI search pipeline, and what the verified evidence says per stage. The plausible channels are index-side (stages 1–2, Google/Bing stack only). The reading stage — where vendor explanations usually place the mechanism — is where the direct evidence is most negative.

The honest per-stage scorecard: at crawl and index time, Google and Bing demonstrably parse structured data, Bing has said on stage that schema helps its LLMs understand content — and Google’s own documentation now states “there’s also no special schema.org structured data that you need to add” for AI features. Google has meanwhile been retiring schema-consuming surfaces, including FAQ rich results in May 2026. At retrieval time, rank dominates and rank’s relationship to schema is unproven — though a February 2026 lab benchmark, SAGEO Arena, offers the sharpest what-if in the literature: in a pipeline built to index structural fields including JSON-LD, enriching those fields lifted retrieval 22%, while the classic body-text GEO tactics actively hurt it. The catch is the conditional. No one has shown a commercial engine indexes markup that way, the study pooled schema with titles and headings rather than isolating it, and even in that markup-friendly pipeline the generator overwhelmingly quoted body text. At reading time, the public evidence says most systems never see the markup at all. And at generation time, models cite based on what survives into their context.

That reading-stage claim has unusually direct evidence. A German technical-SEO firm, searchVIU, planted a product page with prices embedded eight different ways and asked five AI systems to fetch it: none used any hidden markup — page JSON-LD, injected JSON-LD, hidden microdata, hidden RDFa, all ignored, while visible content was extracted fine. The extraction libraries that feed LLM training corpora behave the same way: the FineWeb dataset paper documents that the major open pipelines extract text with trafilatura, a tool that discards script tags — and JSON-LD lives in a script tag.

price placed in…ChatGPTClaudeGeminiPerplexityAI Mode
Visible body text——
Visible microdata / RDFa——
JSON-LD in page (hidden)——
JS-injected JSON-LD (hidden)——
Hidden microdata (hidden)——
Hidden RDFa (hidden)——
The searchVIU fetch experiment (Dec 2025): a planted product page with prices embedded in different layers, fetched live by five AI systems. Filled = value extracted; empty = ignored; dash = system is index-based and returned nothing for a fresh page (Perplexity, AI Mode), or extracted no prices at all (Claude, visible ones included). No system read any hidden-in-markup value. Gemini was the only system to execute JavaScript.

There’s one nuance worth keeping: this isn’t a law of physics, it’s an architectural choice per product. Otterly’s fetch tests found exactly one system — Gemini — that could retrieve a page’s JSON-LD on request. Frontier models read markup perfectly well when a raw page reaches them; most pipelines just don’t let it reach them.

I want to end with the part of this review that unsettled me more than any individual result: watching claims mutate as they travel. The BrightEdge 44% is a ghost. Search Engine Land dated the SearchAtlas study to “December 2024” — the study was published December 14, 2025. An aggregator attributed GEO-16 to “UC Berkeley,” which is half-true at best (the preprint’s authors list Berkeley affiliations alongside a commercial research outfit, and the paper is not peer-reviewed). The KDD generative-engine-optimization paper — which never tested markup at all, only visible-text edits — is routinely cited as schema evidence because its “up to 40% visibility” number is too quotable to leave alone.

"FAQ schema: +44% AI citations (BrightEdge)"
No such study exists. A real BrightEdge 44% stat is about AI Overviews criticizing brands — the number changed hosts.
"A December 2024 study found no correlation"
The matching study was published 14 December 2025. The date shifted a year in trade-press retelling.
"A UC Berkeley study showed +39% from structured data"
A non-peer-reviewed preprint with dual Berkeley/commercial affiliations — measured on a corpus containing only already-cited pages.
"Research shows GEO tactics lift visibility 40% — add schema"
The KDD 2024 GEO paper tested nine visible-text edits. It never tested markup.
Four claims as they circulate (left) versus what the primary sources show (right). Each mutation is small — a transplanted number, a shifted date, an upgraded affiliation, a swapped subject — and each survived until someone fetched the original.

This is the mechanism by which an industry consensus forms without evidence underneath it. A real number from one context gets reattached to a claim from another; a date shifts; an affiliation gets upgraded; a study about content edits becomes a study about markup. Each mutation is small. The compound effect is a field that is very sure of something nobody has demonstrated.

What would actually settle the question is the study nobody has run: a randomized test where the only difference between pages is the markup itself — same visible content, same site, same rank conditions — measured per engine, over enough time, with the statistical corrections Fischman showed are necessary. Fischman’s own paper closes by calling for exactly this. Until someone runs it, the honest summary of the schema-and-AI-citations literature is four sentences: the correlations are real and uncontrolled. The best-controlled studies found no effect — or a small negative one. Any other number you have heard on this topic is worth checking against its original source before you repeat it. And the effort schema was going to absorb belongs where the evidence actually points: toward ranking, and toward concrete facts visible on the page.

How this review was done — and its limits

The review itself is run the way I’d want any study I criticize to be run: documented queries, explicit inclusion criteria, every included claim verified against its primary source, an independent second-reviewer pass that re-screened all 71 papers blind and re-verified a random sample of figures, and a documented trail of what was excluded and why.

120Records identified (arXiv sweep, deep-research, targeted, chased, prior-known)
71Unique arXiv papers title-screened
30Primaries fetched and verified
19Graded evidence entries
10Direct schema ↔ citation studies
The review's screening flow. Five documented arXiv queries (re-run the following day: zero new records), plus an AI deep-research sweep, targeted searches, and citation chasing. Every included figure was verified against its primary source; an independent second-reviewer pass re-screened all 71 papers blind and re-verified a seeded random sample of claims, 8 of 8 reproducing exactly.

It has limits worth stating just as plainly: the protocol was written alongside the search rather than pre-registered; the first screening pass was one reviewer, and a blind second pass is a check, not a cure; a handful of sources were verified at abstract level rather than full text; and everything here is an English-language snapshot taken in early July 2026, in a field that turns over monthly.

Sources

Every figure in this piece was verified against the primary source listed below.