AI cited pages have been nearly thrice extra more likely to have JSON-LD than non-cited pages.
That’s a giant hole, and the sort of stat that will get shared in LinkedIn carousels and convention slides as proof that schema is an AI visibility lever.
However we weren’t happy with the info because it may simply have been correlation, not trigger.
Schema markup tends to stay on better-maintained, extra technically refined websites, and those self same websites publish stronger content material, construct extra authority, earn extra hyperlinks, and do all the opposite issues that get pages cited.
Schema may very well be doing actual work, but it surely may additionally simply be using the wave of each different sign.
So we couldn’t really reply the query SEOs actually care about: if I add schema to my web page, will I get cited extra by AI?
To search out out, we ran a second examine designed to isolate the impact of including schema.
Right here’s what we discovered.
We tracked 1,885 internet pages that added JSON-LD schema between August 2025 and March 2026, matched them in opposition to 4,000 management pages, and measured quotation modifications throughout Google AI Overviews, AI Mode, and ChatGPT.
Including schema produced no main uplift in citations on any platform.
| AI supply | Impact on citations | Verdict |
|---|---|---|
| Google AIO | −4.6% | Small however statistically important decline relative to matched controls; (each teams have been declining collectively, however handled pages fell barely sooner) |
| Google AI Mode | +2.4% | Statistically indistinguishable from zero |
| ChatGPT | +2.2% | Statistically indistinguishable from zero |
These percentages come from our most dependable evaluation (a matched difference-in-differences [DiD] take a look at).
On this take a look at, each AI Mode and ChatGPT handled pages carried out barely higher than management pages on common, however the variations are sufficiently small that they may simply be random noise throughout hundreds of URLs.
AI Overviews confirmed a 4.6% decline, which is small however statistically important relative to matched management pages.
However that isn’t fairly the complete story—we’ll get into that within the subsequent part.
So, total, we are able to’t inform whether or not the schema did a tiny bit of fine or nothing at all.
AI Overview citations on handled pages fell by 4.6% relative to manage pages, and the result’s “statistically important” (the chances of seeing a niche this massive by pure probability are about 1 in 2,500).
However earlier than anybody reads this as “including schema hurts your AI Overview citations”, there are two issues it’s essential bear in thoughts.
- Absolutely the dimension is small. We’re speaking about a mean lack of round 12 each day citations per web page, in a pattern the place most pages have been getting a whole lot.
- Each handled and matched management pages have been already on a steep downward trajectory earlier than schema was added—the sort of decline you’d count on from AI Overviews pulling again from these particular varieties of content material for causes unrelated to schema (e.g. a Google replace altering what will get surfaced, the content material getting stale, or Google not having recrawled the web page not too long ago).
Sidenote.
The way to learn this chart: each strains are anchored to 1.0 at week −1 (the week earlier than schema was added), in order that they all the time begin on the similar level by design. Earlier than remedy, each teams decline collectively. After remedy, handled pages sit barely beneath the matched controls (that is the −4.6% hole).
That stated, if including schema had no impact on citations both approach, we’d count on handled pages and matched controls to say no collectively on the similar price (which is broadly what we see for AI Mode and ChatGPT).
The truth that handled pages declined barely extra suggests schema had a small damaging impact—but it surely may additionally mirror different elements.
We will’t inform which one it’s from this knowledge alone.
Utilizing Model Radar, Xibeijia pulled just a few million URLs cited in AI Overviews.
She then retrieved the HTML historical past from our crawler database, labeled whether or not every URL contained <script sort="utility/ld+json">, and noticed the date that schema presence transitioned from “False” to “True”.
This left her with 1,885 pages that launched JSON-LD between August 2025 and March 2026.
Lastly, to investigate all of that knowledge, she used Agent A, our new AI advertising and marketing agent.
For every web page Xibieijia knew two key dates:
- The final day our crawler checked the web page and located no JSON-LD
- The primary day our crawler detected JSON-LD on the web page
The day a web page added JSON-LD is its remedy date.
Sidenote.
“Remedy” is the usual time period for the second a change is utilized to one thing we’re measuring.
Xibeijia measured what number of occasions every web page was cited by Google AIO, Google AI Mode, and ChatGPT within the 30 days earlier than and 30 days after the remedy date.
The difficult a part of any examine like that is seeing previous noisy knowledge.
Citations throughout all of AI search have been shifting throughout this era; AI Overviews have been contracting, AI Mode was exploding.
If Xibeija had simply performed a easy before-and-after comparability, it might have been measuring the platform development, not the schema impact.
So for every handled URL she picked 3 management URLs (from completely different domains, with comparable pre-period quotation ranges) that had by no means added JSON-LD.
Evaluating two teams of pages that have been getting cited on the similar price earlier than—the place the one predominant distinction was that one group added schema—made it simpler to isolate what schema really did.
We appeared on the knowledge 4 alternative ways to ensure any conclusion held up below scrutiny.
In every take a look at, we requested a barely completely different model of the query: “did schema do something?”
You solely actually consider a discovering when a number of of them agree and, on this case, they do.
Check 1: In contrast common quotation modifications between handled and management pages (a two-sample t-test).
Sidenote.
The way to learn these charts: every bar exhibits what number of pages skilled quotation change after remedy. Proper of zero = gained, left = misplaced. Handled pages in shade, controls in gray. For AI Overviews, just a few outliers (some shedding 400 a day, some gaining 200) dragged the handled common damaging. Strip them out, and handled and management teams look roughly the similar.
Check 2: Ran a difference-in-differences (DiD) evaluation to strip out platform-wide traits. That is the take a look at we belief most, and the supply of the findings on this article.
Sidenote.
The way to learn this chart: every dot exhibits the impact of schema after stripping out platform traits. The bar round a dot exhibits margin of error—if it crosses zero, the outcome may simply be noise. If we simply appeared on the uncooked before-and-after progress of AI Mode, it got here in at +43%, however this evaluation revealed management URLs gained nearly as a lot, that means AI Mode was exploding for everybody. Strip that out and the +43% shrinks to the +2.4% proven right here.
Check 3: Plotted citations week-by-week to test whether or not handled and management pages have been already drifting aside earlier than schema was added (an occasion examine).
Sidenote.
The way to learn this chart: each strains are anchored to 1.0 at week −1, so they begin on the similar level by design. The form is what issues. Handled and management tracked intently earlier than week 0 and rose collectively after, which factors to a platform-wide AI Mode increase slightly than a schema impact.
Check 4: Re-ran the difference-in-differences (DiD) with a symmetrical window that excluded the recrawling interval, to ensure the outcome wasn’t delicate to how we outlined “earlier than” and “after.”
Sidenote.
The way to learn this chart: every platform exhibits two estimates facet by facet, one for every “earlier than” and “after” definition. The bars across the dot present the margin of error. Each estimates land in roughly the identical place for each platform, so the outcome holds no matter how “earlier than” and “after” are outlined.
All 4 assessments informed the identical story: no quotation progress in AI Mode, no quotation progress in ChatGPT, and a small AI Overview decline that’s actual however sufficiently small that we are able to’t definitively pin it on schema.
Essentially the most constant discovering is that not a lot actually modified—schema had no clear optimistic or damaging impact.
Caveat
The place schema would possibly nonetheless matter: pages not but cited by AI
There’s one vital factor it’s essential find out about this knowledge: we studied pages that have been already being cited closely by AI.
Each web page within the dataset had 100+ AI Overview citations in February 2025, earlier than any schema was added.
These pages have been already contained in the consideration set, being crawled and surfaced by LLMs.
If a web page is already getting picked up, our knowledge means that including schema isn’t going to push it greater.
However for pages that aren’t being seen by AI techniques in any respect, schema markup would possibly nonetheless play a job in serving to them get crawled, parsed, or listed within the first place.
Our examine can’t communicate to that instantly, however a current experiment from searchVIU solutions a associated query.
They examined whether or not 5 main AI techniques (ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode) really used schema markup when fetching a web page in real-time.
Spoiler: none of them did. Throughout direct retrieval, each system extracted solely seen HTML content material. JSON-LD, hidden Microdata, and hidden RDFa have been all ignored.
A number of different factors to flag, and a few questions price testing subsequent:
- Pages that add JSON-LD typically change different issues on the similar time (e.g. hyperlinks, content material, technical fixes). We will’t totally separate schema from these sorts of co-occurrences.
- We pooled all schema sorts collectively. Article, FAQ, Product, HowTo, Group. It’s potential some sorts assist greater than others. This can be price digging into.
- We measured 30 days post-treatment. If JSON-LD has a slow-burn impact, a 60- or 90-day window would possibly reveal extra progress.
- We studied JSON-LD—probably the most extensively used schema format. Different codecs exist (Microdata and RDFa), however we haven’t but examined them.
- We solely checked out schema within the web page’s HTML, not schema injected through JavaScript. AI crawlers seem to deal with the 2 in a different way. ¹
- The small AI Overview decline is actual however unexplained. Handled pages dropped about 4.6% greater than matched controls, and we don’t know why. A follow-up examine may have a look at whether or not particular schema sorts or particular content material sorts account for the hole.
Wish to know whether or not schema works on your website particularly? Run a smaller model of this examine your self. Brand Radar can help when it comes to tracking the course of AI citations:
- Pick 5–10 test pages where you plan to add JSON-LD. Ideally pages already getting some AI citations, so you have a baseline (pages with zero citations make it harder to tell whether schema did nothing, or whether the page just wasn’t going to get cited either way). You can check this in the Cited Pages report.
- Pick 5–10 control pages with similar citation levels that you’re not adding schema to. This is what separates “schema did something” from “AI Overviews shifted for everyone that month.”
- Record baseline citations for both groups across AI Overview, AI Mode, and ChatGPT in Brand Radar. Just apply URL filters to isolate those citation numbers.
- Add schema to your test pages and note the date. Don’t change anything else on those pages during the test window.
- Compare both groups after 30 days (or longer if you can). The question is: “did treated pages go up more than control pages did?”
If both groups moved by similar amounts, that’s more to do with the platform trend than the schema.
But if treated pages outperformed controls, that’s a sign schema is having a positive impact on citations.
If you run this on your own pages and get a different result to ours, let us know.
For pages already getting cited by AI, including JSON-LD schema didn’t increase citations on Google AI Mode or ChatGPT, and confirmed solely a small decline in AI Overviews that we are able to’t clearly attribute to schema.
So why are 53% of AI-cited pages working schema?
As a result of the websites that add structured knowledge are inclined to additionally put money into technical search engine optimization, publish authoritative content material, construct hyperlinks, keep their pages, and rank nicely in common search.
AI techniques usually tend to retrieve this sort of content material, so cited pages over-index on all of these indicators without delay. Strip schema out and it’s very possible the remainder of the indicators nonetheless carry the web page by means of to quotation.
When you’re already doing the remainder of the search engine optimization work nicely, JSON-LD isn’t going to be the unlock. And when you’re not, schema by itself in all probability received’t make up for that.
There are nonetheless, in fact, many good causes to make use of JSON-LD schema (wealthy outcomes, voice assistants, information graphs, downstream entity recognition).
But when the one cause you’re including it’s to get extra AI citations on pages which might be already seen, our knowledge doesn’t assist that guess.
