Does Schema Markup Help AI Visibility? An 1,885-Page Study Says Not How You Think
Written by Martin J.You added schema markup to twenty pages this quarter because someone told you it would get you cited in AI Overviews. Did it actually work? Ahrefs ran the largest controlled test on this — and the answer isn't what most SEO advice prepared you for.
Intro
You added schema markup to twenty pages this quarter because someone told you it would get you cited in AI Overviews. Maybe it was a conference talk, maybe a LinkedIn post with a screenshot of impressive-looking JSON-LD. You did the work — Article schema, FAQPage, the whole checklist. Did your AI citations actually move? For most people who ask themselves that question honestly, the answer is: not really, or not enough to explain the hours spent. A team at Ahrefs just ran the largest controlled test anyone has published on this exact question, tracking 1,885 pages that added schema against 4,000 that didn't. What they found should change how you spend your next SEO sprint — but probably not in the direction you're expecting. Here's what almost nobody is measuring: whether the schema markup on your site actually reaches the AI system that might cite you, or whether it dissolves into meaningless character fragments the moment a language model tokenizes your page. Your JSON-LD looks perfectly structured in Google's Rich Results Test. That test was never checking what an LLM does with it. Somewhere between your JSON-LD script tag and the moment ChatGPT decides what to cite, the structure you built is getting flattened into noise for most sites — and the pages winning citations right now are winning for reasons that have nothing to do with your schema effort at all. So what does explain it?
The belief everyone repeats: "add schema, get cited"
The advice has been everywhere for two years: implement structured data, and AI search engines will understand your content well enough to cite it. FAQ schema for AI Overviews. Product schema for shopping assistants. Organization and Article schema as a baseline requirement. It's clean, it's actionable, and agencies love selling it because it's a deliverable you can point to.
The advice isn't baseless. Earlier correlational research found that pages cited by AI Overviews were roughly three times more likely to carry JSON-LD than pages that weren't cited. You've probably seen that statistic somewhere — it traveled fast. What didn't travel with it: correlation isn't causation, and nobody had actually tested what happens when you add schema to a page and watch what changes.
What an 1,885-page controlled test actually found
Ahrefs closed that gap. Between August 2025 and March 2026, its team tracked 1,885 pages that added JSON-LD schema and matched them against 4,000 control pages using a difference-in-differences analysis run through its Brand Radar tool — the closest thing to a real experiment this topic has had. The results, published May 11, 2026:
AI platform | Citation change after adding schema | What it means |
|---|---|---|
Google AI Overviews | −4.6% | Small but statistically real decline — the odds of this happening by chance are roughly 1 in 2,500 |
Google AI Mode | +2.4% | Statistically indistinguishable from zero |
ChatGPT | +2.2% | Statistically indistinguishable from zero |
Read that table again. The one platform where schema showed a measurable effect, it went the wrong way. On the other two, adding schema did nothing that couldn't be explained by noise. If you're using Google AI Overviews specifically as your main citation channel, this is the study you need to know about before you spend another sprint on markup.
There's an honest caveat here, and it matters: every page in the dataset already had 100+ AI Overview citations before schema was added. Ahrefs says this directly — the test can't tell you whether schema helps pages that aren't yet being cited at all. That's a genuinely different, untested population. Hold on to that caveat; it comes back later, in the section on when schema still plausibly helps you.
One more piece from the same research explains why the effect is so weak. A separate experiment — cited in Ahrefs' writeup — had five different AI systems fetch pages directly and checked what they actually extracted. None of them read JSON-LD, Microdata, or RDFa. They pulled from visible HTML text only. If a model never reads your structured data in the first place, it can't be citing you because of it.
Three tests that reached the same conclusion — independently
If this were one study, you could reasonably call it interesting and move on. It isn't just one study. Three more tests, run by different people, using different methods, on different platforms, reached the same conclusion — and two of them predate Ahrefs by about eight months.
The tokenization problem: your schema never reaches the model
In September 2025, SEO practitioner Mark Williams-Cook posted visual proof of what happens to schema inside an LLM's tokenizer. The Organization type declaration doesn't survive as a structured concept — it gets split into separate token fragments, indistinguishable from any other stray text on the page. The graph relationship you built is gone before the model ever "sees" it as a graph.
The schema-only page nobody could read
Around the same time, Julio C. Guevara ran a more direct test: two pages, one with visible text plus schema, one with only structured data and zero visible text. He ran hundreds of extraction prompts against Gemini and ChatGPT, asking for price, color, and SKU. Result: zero retrieval from the schema-only page. If your key information exists solely inside JSON-LD and nowhere in the text a person would actually read, the model can't find it on your page either.
107,352 URLs, one verdict on Google AI Mode
The largest of the three, and the one you should weigh most heavily: salt.agency analyzed schema-type distribution across 107,352 URLs cited in Google AI Mode. The breakdown — Organization on 82%, WebPage/Article on 76%, BreadcrumbList on 59%, FAQPage on 41% — matches almost exactly what any well-built, normal website already has. No schema type showed a measurable citation advantage over that baseline. Dan Taylor, who led the analysis, put it plainly:
Schema is a hygiene factor (at best) for AI Mode visibility, not a differentiator.
Three separate teams, three separate methods, one conclusion. If you're still weighing whether to trust the Ahrefs result alone, you don't have to — it isn't standing alone.
The stat everyone misquotes (and what it actually says)
There's an academic paper behind a lot of "rich schema wins" claims you've probably seen shared without anyone reading past the abstract. K. Fischman's SSRN working paper analyzed 730 AI citations across ChatGPT and Gemini. The naive, pooled version of the result showed schema negatively associated with citation. After correcting for how the control set was built and clustering errors by query, that association collapsed to statistically null.
Here's the number that keeps getting quoted out of context, and you've likely seen it in a deck somewhere: "attribute-rich" schema — Product or Review markup with real prices, ratings, and specs filled in — was cited 61.7% of the time, versus 41.6% for generic schema like plain Article or BreadcrumbList tags. A 20-point gap. That sounds like proof rich schema wins, and people keep repeating it as exactly that.
What almost every summary leaves out: pages with no schema at all sat at 59.8% — statistically indistinguishable from the "winning" 61.7%. The real driver isn't the schema richness. It's that pages with populated prices, ratings, and specs tend to have genuinely richer visible content too, and the schema comes along for that lift without causing it. A related field study of the same 730-citation dataset found that entity-linking sophistication — Wikidata identifiers, the kind of thing schema purists care most about — produced a flat null result. And 63.5% of the AI-cited pages in that dataset didn't even appear in the organic top-10 for the query that surfaced them, which tells you AI citation isn't just a restatement of your Google ranking.
One more misquote you should retire from your deck: the widely-shared "FAQ schema lifts citation 41% vs. 15%" figure comes from a roughly 50-domain vendor study that never ran a controlled before/after test — it's the same correlation-as-causation problem as the original Ahrefs 6-million-URL finding. And Google quietly deprecated FAQ rich results shortly before the 1,885-page study even published, which undercuts the classic "add FAQ schema for AI visibility" advice at its foundation, not just its evidence.
What's actually driving AI citations in 2026
If schema isn't the lever, what should you actually be pulling? From Ahrefs' broader 2026 tracking and independent citation trackers, five signals show up consistently in the pages that do get cited — and the honest framing matters here, because even domain authority is a smaller piece of it than most guides suggest:
- Domain authority is a weaker signal than you'd think. The DA-to-citation correlation has dropped to r=0.18 in 2026 tracking, and 47% of AI Overview citations now come from pages ranking below position #5 in organic search. It still matters — it's just one input, not the top one.
- Content directness beats length. A March 2026 study of 174,048 pages and 560,346 AI Overviews found a Spearman correlation of just 0.04 between word count and citation position — basically zero. Average cited content runs 1,282 words, but 53.4% of cited pages are under 1,000 words. Dense, self-contained answers beat padded ones.
- Verifiable authority signals matter more than you're investing in. 96% of AI Overview citations now come from sources with a named author and linked bio, cited and linked external sources, a visible publication date, and no unsourced statistical claims.
- Freshness is pulling ahead of pure ranking. Top-10 organic results accounted for only 38% of AI Overview citations in March 2026 — down from 76% in July 2025. Ranking well is decreasingly sufficient on its own.
- Earned mentions off your own site count. Reddit is the single most-cited domain across ChatGPT, AI Mode, Gemini, Perplexity, and AI Overviews combined. LinkedIn is close behind and rising fast for B2B queries — citation frequency there roughly doubled between November 2025 and February 2026. Brand mentions on third-party platforms correlate with citation increases within two to four weeks.
For the fuller framework on prioritizing these five signals against each other, see our guide to AI visibility optimization. And if you're tracking this across ChatGPT and Gemini specifically rather than Google's AI surfaces, our LLM visibility pillar breaks down what changes platform by platform.
When schema still plausibly helps — the honest exceptions
None of this means you should rip out your schema tags tonight. Three narrow cases still hold up for you:
- Rich results in classic Google Search. This is a separate benefit from AI citation entirely, and it's still real — star ratings, breadcrumbs, and FAQ dropdowns in regular search results still come from schema.
- Cold-start pages not yet in the AI consideration set. Remember Ahrefs' own caveat — every page in its dataset already had 100+ citations before schema was added. Whether schema helps a brand-new page get discovered and considered in the first place is genuinely untested. Plausible, not proven.
- Baseline hygiene, not a lever. Organization, Article, Person, and BreadcrumbList schema show up on 76–82% of AI Mode citations — but that's because they're present on any well-built site, not because they cause the citation. Keep them. Don't expect them to change your citation numbers.
The actionable takeaway
Don't spend your next sprint on a schema audit expecting it to fix a citation problem. Spend it on the things four independent studies keep pointing at instead: making your first paragraph answer the question directly, naming your authors, citing your sources, keeping content current, and earning mentions on the platforms AI models actually pull from. Schema stays on your checklist as hygiene — implement it once, correctly, and move on. It doesn't stay there as your strategy.
Before you touch another line of JSON-LD, it's worth knowing where you actually stand. Our AEO audit walks through exactly which of these signals your site is missing — which is usually a more productive hour than adding another schema type nobody's model is reading.
See what's actually driving your AI citations
Schema, content depth, authority signals, freshness — tracking all of it by hand across ChatGPT, Gemini, and Google's AI surfaces is its own job. Allable's AI visibility dashboard tracks schema impact alongside the content quality signals that the research above shows actually move citations, so you can see which lever is worth pulling next instead of guessing.
Frequently Asked Questions
- Does schema markup help you get cited in ChatGPT?
- Not measurably, according to the largest controlled test on this to date. Ahrefs' 1,885-page study found a +2.2% change in ChatGPT citations after adding schema — statistically indistinguishable from zero. Google AI Mode showed a similarly null +2.4%.
- Should I remove FAQ schema now that Google deprecated FAQ rich results?
- You don't need to rip it out, but don't add more of it expecting an AI citation boost. Google deprecated FAQ rich results shortly before the Ahrefs study published, and the widely-cited '41% vs. 15%' FAQ schema stat came from an uncontrolled correlational study, not a real before/after test.
- If schema doesn't help AI citations, does it still help SEO at all?
- Yes — that's a separate question from this article's focus. Schema still powers rich results in classic Google Search (star ratings, breadcrumbs, FAQ dropdowns), which is a real, ongoing benefit independent of AI citation behavior.
- What's the actual difference between Google AI Overviews and AI Mode for schema?
- AI Overviews showed a small but statistically real decline (-4.6%) after schema was added in Ahrefs' test. AI Mode and ChatGPT both showed statistically null results. None of the three platforms showed schema producing a meaningful citation lift.
- How do I know if it's my content, not my schema, holding back my AI citations?
- Run an audit against the signals that do correlate with citations: does your first paragraph answer the query directly, is your author named and linked, are your sources cited, is the content recently updated, and are you earning mentions on platforms like Reddit and LinkedIn? Our AEO audit checks all five.