Schema markup won't get you cited by AI — here's what actually will

A 1,885-page Ahrefs study and a live-fetch test across five AI systems both say the same thing: adding schema barely moves AI citations. Here's what does.

If you have spent this year adding JSON-LD to every template in your CMS hoping it would get you quoted inside a ChatGPT answer or a Google AI Overview, you are not alone. It has become one of those pieces of SEO folk wisdom that everyone repeats: pages cited by AI have schema, so schema must be why they get cited. Add the schema, get the citation.

Two separate studies published this year put that assumption through an actual controlled test, and the results are a lot less flattering to structured data than the LinkedIn carousels suggest.

Why this matters right now

Google’s AI Overviews now show up on roughly a quarter of all US searches, and on about half of purely informational queries. A growing share of people never click through to a website at all — they read the AI-generated summary and move on. So the question of what actually gets a page selected, quoted, and linked inside that summary is not academic anymore. It is the difference between organic traffic that still shows up in your analytics and traffic that quietly evaporates into an answer box.

Schema markup sits right in the middle of that anxiety. It is cheap to implement, it is something a developer can ship in an afternoon, and for years it has been sold as one of the more “future-proof” technical SEO investments. So when two independent studies test whether it actually earns you an AI citation and both come back with “not really,” that is worth sitting with.

What the Ahrefs study actually measured

Ahrefs started by pulling a few million URLs cited in Google AI Overviews and checking how many carried JSON-LD schema. The gap was stark — cited pages were almost three times more likely to have schema than pages that were not cited. That is the statistic that usually ends the conversation and gets turned into a slide.

But a correlation like that is exactly the kind of thing that falls apart under a real experiment, because sites that bother installing schema also tend to be the sites doing everything else right: stronger content, better technical SEO, more backlinks, more consistent publishing. Schema could be doing real work, or it could just be riding along with every other ranking signal a mature site already has.

To separate the two, Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matched each one against three control pages with a similar citation history that never added schema, and measured what happened to citations across Google AI Overviews, Google AI Mode, and ChatGPT in the following month. Because both groups were already being pulled into AI answers at a similar rate before the change, any gap that opened up afterward could reasonably be pinned on the schema itself rather than on general momentum.

Four different statistical tests, run four different ways to make sure the result wasn’t an artifact of how the before-and-after window was drawn, all landed in the same place: adding schema produced no meaningful citation lift on any platform. AI Mode and ChatGPT citations moved by roughly plus 2%, which is statistically indistinguishable from noise. AI Overviews actually dipped by 4.6% relative to the control group — small in absolute terms (around 12 fewer daily citations on pages that were getting hundreds), and Ahrefs is upfront that they cannot fully explain the dip, but it is the opposite of the uplift schema is supposed to deliver.

One caveat worth keeping: every page in that dataset already had over a hundred AI Overview citations before the test began. These were pages already inside the “consideration set,” already being crawled and surfaced regularly. The study cannot tell you what schema does for a page that AI systems are not looking at yet — only that, for pages already on the radar, schema is not the lever that pushes them higher.

What a live-fetch test found about what AI actually reads

The second piece of the puzzle comes from a much smaller but more mechanically revealing experiment. SearchVIU built a test page listing several fictional products, split the pricing information across different sources — plain visible HTML, JavaScript-rendered content, hidden JSON-LD, hidden Microdata, and hidden RDFa — then asked ChatGPT, Claude, Gemini, and Perplexity to report the prices.

The result: not one of the five systems tested extracted a single price that existed only in JSON-LD schema. Not Gemini, which was otherwise the strongest performer and the only one to execute JavaScript during a live fetch. Not ChatGPT, which correctly read prices sitting in plain visible HTML. During direct retrieval, every system parsed visible content on the page and ignored the structured data markup entirely, treating it as if it were not there.

That does not mean schema is worthless everywhere in the pipeline — it very plausibly still gets extracted during the indexing phase, which is where Google’s AI Overviews and Bing’s Copilot draw from, since they are querying a search index rather than fetching your page live in the moment. But for the growing category of AI systems that fetch a page directly when a user asks a question, the structured data in your <script type="application/ld+json"> block might as well not exist.

Put the two studies together and a consistent picture forms: schema is not the thing separating a cited page from an uncited one. The pages that get cited are the ones with clear, well-organized, human-readable content that both a person and a parser can understand without needing a hidden data layer to translate it.

What this means for how you build

For a studio that ships static sites, this lines up with a bias we already have. A site where the content that matters is rendered directly into HTML — not injected by JavaScript after the fact, not hidden inside a schema block as the only source of truth — is legible to every one of these systems by default, no extra optimization layer required. → Read also: Why this site is static covers the same argument from the performance and maintenance side; AI legibility turns out to be another reason to keep content in HTML rather than behind a render step.

Here is what actually seems worth prioritizing, based on both studies:

Write the answer in the content, not just the markup. If a fact only exists inside JSON-LD, treat it as invisible to most AI systems today. Put the price, the date, the author, the key claim in a sentence a person would actually read.

Keep using schema for what it was built for. Rich results, Google’s understanding of your entities, voice assistant answers, and knowledge graph connections are all still real reasons to mark up your pages — the Ahrefs data just says “more AI citations” isn’t one of them for pages already getting cited.

Invest the time you would have spent on marginal schema types into earning the actual signals that separate cited pages from the rest: clear, extractable answers near the top of a page, content that stays current, and the backlinks and authority that got those control-group pages onto AI’s radar in the first place.

If you are testing a new page type, do not assume schema is doing the lifting silently. Ahrefs published a lightweight version of their methodology — pick a handful of test pages, hold a matched set back without schema, and check your own citation counts after 30 days. It is the only way to know if your specific case is the exception.

The mistake worth avoiding

The easy overcorrection here is to conclude schema is dead and rip it out. That is not what either study says. What they say is narrower and more useful: schema is not an AI-citation strategy on its own, and the 53% of AI-cited pages carrying schema are cited despite a lot of other things being true about those sites, not because of the markup.

If your roadmap this quarter has “add more schema types” sitting above “make our actual content easier to lift a clean answer out of,” this is a reasonable moment to flip that order.

FAQ

Does this mean I should remove schema from my site? No. Neither study measured any downside to schema outside that one narrow 4.6% AI Overview dip, and schema still does real work for rich results, entity understanding, and voice assistants. The finding is about a specific claim — “schema gets you cited by AI” — not about schema as a whole.

Which schema types were tested? Ahrefs pooled all types together (Article, FAQ, Product, HowTo, Organization, and more), so the study can’t tell you whether one type outperforms another. That is an open question the researchers flagged for follow-up work, not something either study rules out.

Does this apply to pages that have never been cited by AI before? Not directly. Ahrefs’ dataset only included pages that already had over 100 AI Overview citations before the test — pages already inside the “consideration set.” Whether schema helps a brand-new or previously invisible page get picked up in the first place is a separate question neither study answers.

What about schema injected via JavaScript instead of static HTML? Both studies suggest that’s worse, not better. SearchVIU’s test showed AI systems ignore schema regardless of whether it is static or JavaScript-rendered during a direct fetch, and JavaScript-rendered content in general depends on a crawler choosing to execute it — something only one of the five systems tested reliably did.

Sources

{
  "@context": "https://schema.org",
  "@type": "BlogPosting",
  "headline": "Schema markup won't get you cited by AI — here's what actually will",
  "description": "A 1,885-page Ahrefs study and a live-fetch test across five AI systems both say the same thing: adding schema barely moves AI citations. Here's what does.",
  "datePublished": "2026-08-30",
  "dateModified": "2026-08-30",
  "keywords": "seo, ai-search, structured-data"
}