Everyone selling "get cited by ChatGPT" advice puts structured data near the top of the list. I believed it, budgeted work for it, and then read the one study design that can actually test it.
It doesn't hold. I cut that half of my plan.
Let me be precise about the claim, because this is the kind of post that attracts "you clearly don't understand schema" — and on a wider claim, that criticism would be correct.
The claim: adding JSON-LD to a page does not measurably increase how often AI systems cite it.
Not the claim: schema is useless. Rich results in classic search are a different mechanism, they're real, and this study says nothing about them. If you're marking up recipes, products, or events to get carousels and stars in Google, keep doing it — that's a documented feature and it works.
The correlation everyone quotes
The number in circulation comes from an analysis of ~6M URLs: pages cited by AI systems were almost three times more likely to carry JSON-LD.
That's a real finding and it's easy to see why it converted so many people. It's also, per the analysts who published it, confounded — and they say so themselves. Schema markup tends to live on well-resourced, technically competent, established sites. Those sites get cited more for about nine reasons that have nothing to do with markup. The 3× is a property of who ships schema, not of schema.
Which is the classic shape of this mistake: an observational correlation, an obvious causal story, and a whole practice built on it.
The study that can actually answer it
To get causality you need pages that changed, and controls that didn't.
- 1,885 pages that added JSON-LD over an eight-month window
- each matched to 3 control URLs on different domains at similar pre-period citation levels — about 4,000 controls
- 30-day pre/post window
- four statistical approaches: t-tests, difference-in-differences, event-study weekly plots, and alternative-window robustness checks
Matching on pre-period citation level is the part that makes it worth reading. It's what separates "pages that added schema" from "the kind of site that adds schema."
The result:
| Platform | Citation change after adding JSON-LD |
|---|---|
| Google AI Overviews | −4.6% — small but statistically significant decline |
| Google AI Mode | +2.4% — statistically indistinguishable from zero |
| ChatGPT | +2.2% — statistically indistinguishable from zero |
The study's own summary: adding schema produced no major uplift in citations on any platform.
I want to be careful with that −4.6%. It's one platform, one window, a small effect. I don't have a mechanism for it and I'm not going to invent one. The defensible reading is: no uplift anywhere, and possibly a small negative in AI Overviews. Anyone telling you schema hurts AI citation is over-reading this as badly as the people telling you it triples them.
Why it's plausible that markup does nothing here
It's worth asking why we expected it to work.
Schema was designed so a parser could extract typed facts from an HTML document without understanding language. That was genuinely hard in 2011.
An LLM reads the prose. It does not need FAQPage to notice that a paragraph answers a question, any more than you do. The extraction problem schema exists to solve is the problem these models incidentally solved. So the prior should have been "probably no effect," and the burden should have been on the uplift claim from the start.
Retrieval and ranking are a separate question — but that's about whether your page is found, and JSON-LD isn't what gets you found.
What I did instead
I dropped FAQPage/Article markup as a citation tactic. I kept exactly one thing from that half of the plan: an extractable definition in the first paragraph of every explainer page.
Not because a crawler needs it. Because a human landing from a search wants the answer above the fold, and a model summarizing the page will quote the clearest available sentence. That's a content decision with an independent reason to exist, which is the test I now apply to anything with "AEO" attached to it: if the AI angle evaporated tomorrow, would I still do this?
Definition-first paragraph: yes. Hand-maintaining JSON-LD across every page for citation share: no.
(I still ship schema where it does a job — Dataset on an open-data page, so it can appear in dataset search. That's a documented indexing feature, not a citation bet.)
The general lesson
Every emerging channel grows a tactics industry faster than it grows an evidence base, and the tactics that spread are the ones that are legible — a thing you can add to a page and check off. "Write something worth citing" doesn't fit in a checklist, so it loses to "add this JSON block," even when the JSON block is inert.
Two questions that would have saved me the detour:
- Is the evidence causal or correlational? If a claim rests on "cited pages are N× more likely to have X," ask what else is true of sites that ship X.
- Would I do this if the trend disappeared? If no, you're not doing engineering, you're buying a lottery ticket with your afternoons.
If someone has a matched study pointing the other way, I'll change my mind again — post it in the comments and I'll read it. That's roughly the only thing that will.
What I do publish, and why, is on the method page.
Top comments (2)
i tried stripping out the JSON
We need to produce a short YouTube comment per developer style. Must be specific reaction or question about this video. Should be casual, lower case start, one or two sentences, possibly fragment. No quotes, no hashtags, no markdown. Avoid formal language. Must not include URLs. No marketing. Let's craft: "i tried removing json-ld from my aeo plan and saw a drop in click-throughs, anyone else notice the same?" That works. Ensure no em dashes. Use straight quotes only if needed but we