Schema markup doesn’t move AI citations. Stop pretending.

Ahrefs ran the test that the GEO industry has been quietly avoiding for two years. They tracked 1,885 pages that added JSON-LD schema, matched each one against control pages from different domains, and measured citation changes across Google AI Overviews, AI Mode, and ChatGPT over a 30-day window.

The result: nothing. AI Overviews dropped 4.6%. AI Mode rose 2.4%. ChatGPT rose 2.2%. Three of those numbers are statistically indistinguishable from noise. The fourth is a small negative.

This is the cleanest controlled test we've seen on a tactic that's been sold as essential for AI visibility since the moment AI visibility became a thing people would pay to fix. And the data says adding schema to a page that's already in the AI consideration set does not increase citations.

I want to be careful about what this means and what it doesn't.

What the test actually rules out

The Ahrefs methodology is the right one. Difference-in-differences with matched controls is how you isolate the effect of a single change from everything else going on. Pages that add schema also tend to be on sites that invest in content, earn links, fix technical debt, and generally do the work. That's why the correlation between schema presence and AI citations looked so strong in the first place — pages with schema are three times more likely to be cited.

But correlation collapses when you control for site quality. The schema isn't doing the lifting. The everything-else is.

There's also a searchVIU experiment cited in the report that's worth pausing on. They tested whether five AI systems used schema markup when fetching pages in real time. None did. They extracted visible HTML and ignored JSON-LD, Microdata, and RDFa entirely.

That doesn't prove schema plays no role in training or indexing — those are different pipelines. But it does undercut the cleanest version of the GEO pitch: that AI systems read your structured data when answering a query and reward you for having it.

They don't. They read your text.

What the test doesn't rule out

The dataset is pages that were already cited at least 100 times. Every page in the sample was already visible to AI. So the test answers one question — *does adding schema to an already-visible page increase citations?* — and not the question of whether schema helps a page get into the consideration set in the first place.

That second question is harder to test and Ahrefs is honest about not having answered it. Schema may still help crawling, parsing, and entity disambiguation for pages that are climbing the eligibility curve. It just doesn't act as a multiplier once you're already there.

I'd also note that the study pooled all schema types together. Product schema, Article schema, FAQ schema, HowTo, Organization — all in one bucket. It's possible some types matter and others don't, and pooling them hides the signal. But that's a hypothesis the data can't currently support either way.

Why this matters more than the report lets on

The GEO consulting market has been running on the assumption that schema is a primary lever. Tools have been built around it. Audits have been priced around it. Entire agency service lines exist to retrofit JSON-LD across enterprise sites with the explicit promise that it will improve AI search performance.

The honest GEO playbook keeps looking more and more like the honest SEO playbook with a slightly different vocabulary.

If the strongest controlled test we have says the effect is zero or negative, that's a problem for a meaningful chunk of the industry. Not because schema is useless — it still drives rich results in traditional search, still feeds knowledge graphs, still helps with entity recognition — but because it's been mis-sold as the AI visibility lever when it's actually a traditional SEO hygiene item.

The honest GEO playbook keeps looking more and more like the honest SEO playbook with a slightly different vocabulary.

Mike King's work on the 499 status code, which I wrote about last week, points in the same direction. Page reliability and fetch speed determine whether your content survives the retrieval step. Profound's data showed pages with high failure rates received 18x fewer citation events. Those are the levers that actually gate eligibility — not whether you wrapped your content in the right JSON-LD.

What this means for the work

If you're a consultant, the question to ask before recommending a schema project is whether you'd recommend it in the absence of any AI search consideration. If the answer is yes — because rich snippets, knowledge panels, or entity clarity matter for the client — do it on those grounds. If the answer is no and the only justification is "AI visibility," you're selling something the data doesn't support.

If you're in-house and an agency is pitching you on a structured data overhaul as the centrepiece of your AI search strategy, ask for their controlled evidence. Not correlation. Not case studies where six things changed at once. A real before-and-after on isolated schema deployment. They won't have it, because nobody had it until last week, and what now exists points the wrong way.

The actually high-leverage things — content quality, brand strength, fetch reliability, internal linking, earned mentions on sites AI systems trust — are the same things that moved traditional rankings. The mechanism has changed. The signals haven't. That's been my position since AI search became a category, and the evidence keeps lining up behind it.

The schema cargo cult was the easiest part of the GEO pitch to sell because it's the most tactical. It feels like a thing you can do. You can audit it, deploy it, screenshot the Rich Results Test, and put a number on the slide. That's why it became the centrepiece. It's also why it should have been suspicious from the start. The things that genuinely move discovery are rarely the things that audit well in a single afternoon.

Adding JSON-LD to a page that AI already cites does roughly nothing. Sometimes the most useful research is the kind that tells you to stop doing something.