- A matched, controlled Ahrefs study of 1,885 pages that added JSON-LD schema found the effect on AI citations statistically indistinguishable from zero on ChatGPT and Google AI Mode, and a small but statistically significant decline on Google AI Overviews, pooled across every schema type.
- A separate working paper that isolated schema type found a real exception: Product/Review schema populated with actual pricing, ratings, or specs was cited at 61.7%, versus 41.6% for generic schema types (p = .012, 730 citations analyzed).
- The two studies aren't in conflict. Ahrefs pooled all types and explicitly flagged that as an open question; the type-level study is effectively the follow-up, and a peer-reviewed SIGIR 2026 study (252,000 trials) independently confirms the mechanism: it's the underlying data (price, freshness, relevance) driving citation, not the markup format encoding it.
- Ahrefs is also explicit about a real limitation: every page in their dataset already had 100+ AI Overview citations before schema was added. Neither study says whether schema helps a brand-new, not-yet-visible page get into the citation pool in the first place.
- Schema markup still does real work outside of AI citation specifically (rich results, Knowledge Graph feeding, entity disambiguation), but as a Stage 3 citation lever, it only earns its keep when it's carrying data worth citing.
AI visibility strategy already flags schema markup as one of the tactics with close to zero measured effect on AI citations, alongside llms.txt. This piece is the full version: what the actual controlled study found, the real exception a follow-up study found once it isolated schema type, and what a peer-reviewed study says is actually driving both results.
What did the controlled test actually find?
Ahrefs identified 1,885 pages that added JSON-LD schema markup between August 2025 and March 2026, then matched each one against three control pages from different domains with similar citation levels before the schema was added: 4,000 control pages in total. They measured citation counts across Google AI Overviews, Google AI Mode, and ChatGPT for 30 days before and 30 days after the change, using four separate statistical approaches (difference-in-differences, two-sample t-tests, event-study analysis, and sensitivity testing) rather than a single pass.
The results, by platform:
| Platform | Effect after adding schema | Assessment |
|---|---|---|
| Google AI Overviews | -4.6% | Statistically significant decline; odds of a gap this size by chance, about 1 in 2,500 |
| Google AI Mode | +2.4% | Statistically indistinguishable from zero |
| ChatGPT | +2.2% | Statistically indistinguishable from zero |
The AI Overviews decline is real in a statistical sense, but small in absolute terms: an average loss of roughly 12 daily citations per page, on pages that were mostly getting hundreds. Ahrefs' own framing: schema markup did not produce a measurable citation gain on any platform tested, and produced a small measurable loss on one.
What correlation makes this confusing?
Before running the controlled test, Ahrefs' broader analysis of 6 million URLs found something that looks like the opposite conclusion: AI-cited pages were almost three times more likely to carry JSON-LD schema than non-cited pages, and 53% of AI-cited pages in that dataset had schema markup at all. That's a real, measured correlation, and it's exactly the kind of number that gets lifted into a "schema helps AI citations" claim without the controlled test attached to it.
The controlled test is what separates the two. If schema markup itself were driving citations, adding it to a matched set of pages should have produced a measurable gain. It didn't. The likelier explanation, consistent with both findings: schema markup is common on sites that are already authoritative, well-maintained, and doing the things that do drive citations, which makes schema a marker of a well-run site, not a cause of being cited.
Ahrefs is upfront about one more limitation worth carrying forward: they pooled all schema types together (Article, FAQ, Product, HowTo, Organization) rather than isolating which type was added, and note directly that "it's possible some types help more than others... this may be worth digging into." That's a real open question their own study doesn't resolve.
What exception does the pooled data hide?
A working paper by Kurt Fischman, posted to SSRN in February 2026, is effectively that follow-up. Fischman collected 730 AI citations from ChatGPT (GPT-4o with web browsing) and Gemini (1.5 Pro with search grounding) across 75 commercial queries in five categories (SaaS/Technology, Health/Medical, Finance/Insurance, Professional Services, and How-To/DIY) against a control set of 1,006 pages pulled from Google's top-10 organic results for the same queries.
Isolating by type rather than pooling everything together, the study found pages carrying Product or Review schema populated with real, concrete fields (pricing, aggregateRating, specifications) were cited at 61.7%, versus 41.6% for pages using generic schema types like Article or Organization. That's a real, statistically significant gap (p = .012), not noise.
One detail worth including precisely because it shows the study checking its own work: Fischman's initial pooled analysis actually showed a spurious negative association between schema and citation, the opposite of the finding above. Tracing it down, the cause was a methodological artifact: Google's own ranking algorithm systematically enriches top-10 organic results for schema-bearing pages, which inflated schema's prevalence in the non-cited control population specifically and made schema look harmful before the confound was corrected for. That's exactly the kind of check that separates a study worth citing from a stat that just sounds plausible.
This isn't a contradiction of the Ahrefs result. It's a resolution of the exact open question Ahrefs' own pooled analysis left dangling. No net effect across every schema type, pooled, is entirely consistent with a real effect for one specific type (attribute-rich Product/Review schema) getting averaged out by five other types that carry no comparable data and show no comparable lift.
What does a separate, peer-reviewed study say actually matters?
"What Gets Cited: Competitive GEO in AI Answer Engines" (Vishwakarma, Kumar & Jamidar), presented at SIGIR '26 (the 49th ACM SIGIR Conference), ran 252,000 controlled trials across six language models, using a two-document retrieval testbed: two candidate sources injected per trial, paired comparisons across 18 content factors, measuring which source got cited first.
Their finding: topical relevance and list position are the biggest drivers of being cited first. Including explicit price information and a recent timestamp helps consistently. Completeness and trust cues add smaller gains. Formatting-only edits (which is what generic schema markup fundamentally is, a way of formatting data that could just as easily appear in plain text) had little impact on their own.
That's the mechanism both other findings are consistent with, not competing with. A page with a clearly stated price, in schema or in prose, has the underlying data a model can extract and cite. A page with the identical price wrapped in JSON-LD but nowhere meaningful for a generic schema type to point at isn't necessarily better off. Product and Review schema shows a real lift specifically because that schema type is built to carry exactly the data (price, rating, specs) the SIGIR study independently identifies as a gatekeeper. The wrapper earns its keep only when there's a gatekeeper fact inside it.
What does this mean practically?
Don't treat schema markup broadly as a to-do item for AI visibility. Generic and formatting-only schema types show no measured citation benefit on their own. Product and Review schema, specifically populated with real pricing, ratings, or specifications, does show a real, statistically significant one, not because the tag exists, but because it guarantees the exact data a model is looking for is present and structured. If a page genuinely has a price, a rating, or a spec sheet, marking it up with the matching schema type is worth doing. Marking up an "Article" or "Organization" page with schema for its own sake isn't.
Schema also still does real work unrelated to AI citation specifically: rich-result eligibility in classic Google search, and reinforcing entity data for Knowledge Graph resolution (the actual job schema does well, and the reason entity consistency work still benefits from it). Those are real, separate reasons to have it regardless of the citation question. But if the goal is specifically getting cited in a generated AI answer, the SIGIR study's gatekeepers (topical relevance, position, explicit pricing, freshness) are where the actual effort belongs. State the price. Update the timestamp. Answer the actual question the page is about, near the top. Wrap that data in the matching schema type afterward if it exists; the wrapper alone was never what earned the citation.
Frequently asked questions
Does schema markup help you get cited by AI models?
Pooled across all schema types, no. Ahrefs matched 1,885 pages that added JSON-LD schema between August 2025 and March 2026 against 4,000 similar control pages and found the effect on ChatGPT (+2.2%) and Google AI Mode (+2.4%) statistically indistinguishable from zero, with a small but statistically significant decline on Google AI Overviews (-4.6%). But a separate study that isolated schema type found a real exception.
What's the exception?
Product and Review schema populated with real, concrete fields: pricing, aggregateRating, specifications. A working paper by Kurt Fischman (SSRN, February 2026; 730 AI citations from ChatGPT and Gemini across 75 commercial queries, against a 1,006-page control set) found pages with attribute-rich Product/Review schema cited at 61.7%, versus 41.6% for generic schema types like Article or Organization, a statistically significant gap (p = .012). Ahrefs' own study pooled every schema type together and explicitly flagged that as an open question its data couldn't answer; Fischman's study is effectively that follow-up.
So is it the schema or the data driving the citation?
The data. A peer-reviewed SIGIR 2026 study (Vishwakarma, Kumar & Jamidar; 252,000 controlled trials across six models) found topical relevance and list position are the biggest drivers of being cited first, with explicit pricing and a recent timestamp helping consistently, and formatting-only edits having little impact on their own. Fischman's own analysis backs this directly: his initial pooled numbers showed a spurious negative association with schema, traced to Google's ranking algorithm over-representing schema-bearing pages in the non-cited control set itself, a confound his design corrected for. Once corrected, the lift tracked the presence of real data, not the schema wrapper around it.
Does this mean schema markup is worthless?
No, but it means being specific about which schema and why. Generic or formatting-only schema shows no measured citation benefit, per Ahrefs' pooled result. Product/Review schema carrying real pricing, ratings, or specs shows a real one, per Fischman, because it's a reliable, structured way to guarantee that exact data exists on the page. Schema also still does real jobs unrelated to AI citation specifically: rich-result eligibility in classic search, feeding Google's Knowledge Graph, and reinforcing entity disambiguation (Stage 2 of this site's framework).
Sources referenced
- Ahrefs: We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved. (May 2026)
- Fischman, "Does Schema Markup Predict AI Citation? A Cross-Platform Empirical Study of Structured Data and Generative Engine Optimization" (SSRN working paper, Feb 2026, 730 citations analyzed)
- Vishwakarma, Kumar & Jamidar, "What Gets Cited: Competitive GEO in AI Answer Engines" (SIGIR '26, 252,000 controlled trials)
This piece is the full version of a finding already cited in AI visibility strategy and the five stages for winning in the AI search era. For the broader set of extractability patterns that do have measured evidence behind them, see how to write content that AI models can actually quote. For a type-by-type breakdown of which schema markup actually earns its keep, see structured data for AI search.
