Check this page with an assistantOpens a chat asking it to summarise this article and name the evidence behind each claim.
Claude opens with the prompt on your clipboard: Anthropic does not support prefilled prompts on the web, and we would rather copy it than ship a button that drops it.
What the Study Actually Did
Ahrefs published research in May 2026 tracking 1,885 pages that newly added JSON-LD structured data between August 2025 and March 2026. Each treated page was matched against control pages from different domains with similar citation histories that never added schema, and citations were measured thirty days either side of the change.
The design matters more than the headline. This is a matched difference-in-differences test with an event study alongside it, which is a genuinely stronger instrument than the correlation studies that dominate this field. Correlation studies find that cited pages tend to have schema; a matched design asks what happened to pages that added it, against comparable pages that did not.
The results across the three surfaces were small and inconsistent in direction: AI Overviews slightly negative, AI Mode and ChatGPT slightly positive, with the positive movements small enough to be noise. Nothing resembling the effect that structured data services are sold on appeared in any of them.
What This Does Not Show
The finding is narrow and it is being over-read in both directions. It does not show that structured data is worthless, that Google ignores it, or that you should remove it. It shows that adding it did not measurably change how often those pages were cited in AI answers.
| Claim | Status after this study | |
|---|---|---|
| Schema earns rich results where Google supports them | Unaffected and still true, per structured data in 2026 | |
| Product data feeds shopping surfaces | Unaffected, per Merchant Center | |
| Markup helps resolve which entity a page refers to | Mechanistically sound, no measured citation effect, per entity SEO | |
| Adding schema increases AI citations | Not supported by this test | |
| Schema is required to appear in AI answers | Never true; Google states no special markup is needed |
The last row is worth stating separately because it predates the study. Google's own documentation says no special machine-readable files or markup are needed to appear in its search features, including the AI ones. The study is corroboration of something the platform already said plainly and the market kept selling around.
Why This Claim Was So Sellable
Structured data has the three properties that make a deliverable durable regardless of whether it works: it is plausible, it is cheap to produce, and its failure is invisible to the buyer. Nobody can tell whether the schema they paid for did anything.
The mechanism is genuinely plausible, which is what makes it persuasive. Machine-readable descriptions of your content ought to help a machine understand your content. That reasoning is sound and it turns out not to produce the effect, which is exactly why mechanism alone is insufficient evidence and why controlled tests matter.
It is the same shape as the file examined in our review of llms.txt: a reasonable idea, an easy invoice, and no measurable consequence. When a service has those properties, ask for the evidence before the rationale, because a rationale is always available.
What to Do With Structured Data Now
Implement the types that still produce something observable, keep them accurate, and stop treating markup as an AI visibility lever. The whole job for most sites is a day of template work rather than an ongoing service.
- Implement what still renders.
Organization, LocalBusiness, Product, Article, Event, and BreadcrumbList where they genuinely apply. The current list of what earns a visible result is in structured data in 2026.
- Keep it describing the visible page.
Markup asserting things a visitor cannot see is a policy problem rather than a shortcut, and it is checked against the rendered page.
- Do not remove what you have.
Unused structured data does not cause problems, and stripping it is a migration with risk and no upside. The finding is that adding it did not help, not that having it hurts.
- Reprice it in any proposal.
Schema implementation is a fixed piece of template work. If it appears as a recurring line item justified by AI visibility, that justification no longer has evidence behind it.
- Spend the freed budget on what does move.
Citation research consistently points at third-party sources rather than your own pages, which is where the work is, per earning third-party citations.
The Wider Lesson About Evidence
This is the rare case where a specific, monetised SEO claim got a controlled test rather than a correlation study. The result was null. That should recalibrate how much weight mechanism-based arguments carry in this field generally.
Almost every AI visibility tactic currently on sale rests on a plausible mechanism and no measurement. Some of them will turn out to work. The honest position is that we do not know, and the vendors selling them do not know either, which is a materially different claim from the one usually made.
The rule worth carrying forward: never bridge an inspection to an outcome without evidence for the bridge. Observing that your entity data is inconsistent is a fact. Concluding that it is why you are under-cited is a causal claim, and this study is what happens when one of those claims finally gets tested. The measurement discipline that follows is in the evidence ladder.
Questions About Schema and AI Citations
- Does adding schema markup get you cited by AI more often?
A controlled study published by Ahrefs in May 2026 found no meaningful effect. It tracked 1,885 pages that newly added JSON-LD between August 2025 and March 2026, matched each against control pages with similar citation histories that never added schema, and measured citations across AI Overviews, AI Mode, and ChatGPT. The differences were within noise.
- Does that mean schema markup is useless?
No, and this is where the finding gets misread. Structured data still earns rich results where Google supports them, still feeds shopping and product surfaces, and still helps systems resolve which entity a page refers to. What it does not appear to do is increase how often you get cited in AI answers, which is the specific claim being sold hardest.
- How was the study designed?
As a matched difference-in-differences test. Each treated page was paired with control pages from other domains with similar citation histories, and citations were compared thirty days before and after schema was added. The authors ran several analyses including an event study, which is a considerably stronger design than the correlation studies usually cited in this field.
- Why do so many sources claim schema helps AI visibility?
Because it is plausible, easy to sell, and impossible for a buyer to disprove. Markup is a clean deliverable with a fixed price and no observable failure mode. That combination has produced several durable claims in this field, and this one now has a controlled test against it rather than merely an absence of evidence.
- So what should I do about structured data?
Implement the types that still produce something visible, keep it accurate, and stop paying for it as an AI visibility service. The honest sentence is that entity and schema work is hygiene with a sound mechanistic rationale and no measured citation effect, which is a defensible thing to say and a poor thing to invoice against.
