Schema App argues the studies finding no link between markup and AI citations measured the wrong thing. We fetched 20 homepages to see how many sites even publish the kind of markup the argument is about.
Table of contents
- What this is about
- What changed
- What we measured
- Where it matters and where it does not
- How to verify on your own site
- The numbers
- FAQ
Key takeaways
- Google’s documentation is explicit: “There’s also no special schema.org structured data that you need to add” to appear in AI Overviews or AI Mode.
- Of 20 SEO publications we fetched on 7 September 2026, 12 served JSON-LD on the homepage and 8 served none at all.
- Among those 12, 10 used at least one
@idreference that resolved to an entity declared on the same page. Connected markup is the norm where markup exists, not a rarity. - The largest graph belonged to Schema App itself: 28 declared entities and 19
sameAsvalues on one page. - We measured how common connected markup is, not whether it earns citations. Nothing here tests the effect.
What this is about
Schema App published a rebuttal to studies that found no correlation between structured data and AI citations. Their argument: those studies tested basic page-level markup, not a connected entity graph, so they measured the wrong thing. We checked how common each kind actually is. For most sites, no action follows.
What changed
Nothing in the documentation changed. What changed is the argument around it.
Google’s developer documentation on AI features states the requirement plainly:
There are no additional requirements to appear in AI Overviews or AI Mode.
Eligibility, in the same document, is defined as being indexed and able to show with a snippet in ordinary Search. Nothing beyond that.
And, in the same document:
There’s also no special schema.org structured data that you need to add.
Against that, Schema App’s position is that the studies showing no effect are testing the wrong variable. In their words, there is “a significant (and strategic) difference between adding basic Schema Markup to a webpage and using advanced, connected Schema Markup to define your organization’s entities and relationships in a Content Knowledge Graph that lives on your website.”
The distinction they draw is real and it is technical, and it sits inside the wider question of what structured data still earns you. Basic markup describes what is on one page: this is a Product, this is a Recipe. Connected markup declares entities with @id values and then references those identifiers from other objects, so a physician resolves to a specialty, the specialty to a procedure, the procedure to the locations offering it. External references through sameAs point the same entity at Wikidata or another authority.
Their evidence is entirely first-party: a 19.72% rise in AI Overview visibility on their own site, a further 32% quarter over quarter, a customer moving from 27.5% to 36% market share of tracked keywords with AI Overview citations, and 29% and 39% figures from another customer. No control group is described for any of them, and one customer statement that no other SEO work was running at the time is the only nod to isolating the variable. Read them as case studies, which is what they are.
The counter-evidence is stronger on design. As we covered when the 2026 citation data landed, Ahrefs compared 1,885 pages that added JSON-LD against 4,000 matched control pages and found no statistically significant uplift in AI citations, with a 4.6% fall in AI Overview citations as the only significant result. That study has a control group and it is the one Schema App is answering.
What we measured
We took the 20 SEO and search publications in our source list, fetched each homepage once with curl on 7 September 2026, parsed every application/ld+json block, and counted four things: entities declared with an @id, references to an @id, references that resolve to an entity declared on the same page, and sameAs values.
Twelve of the 20 served JSON-LD. Eight served none at all — including Ahrefs, Semrush, Seer Interactive and Search Engine Roundtable, all of which publish research about structured data.
Of the 12 that served markup, 10 used at least one @id reference that resolved on the page. By the definition in the argument above, that is connected markup, and it is what most of these sites already do. Ten sites carried sameAs values. Two sites — Google Search Central and SearchPilot — served markup with no resolving references at all, which is the flat pattern the studies are accused of testing.
The largest graph on any homepage belonged to Schema App: 28 declared entities and 19 sameAs values, though only one internal reference resolved on that page. Their linking is outward, to authorities, more than inward between their own entities.
The spread underneath the headline number is wide. Of the 12 sites with markup, nine carried between five and ten distinct schema types and two carried four or fewer. Every reference we found resolved: across the 10 sites that used references at all, there was not a single dangling edge pointing at an identifier declared nowhere on the page.
The shapes still differ. iPullRank declared six identifiers and referenced all six; Moz declared two and referenced them four times over. Search Engine Journal carried 14 sameAs values against three declared entities, which is heavy outward linking on a thin internal graph. None of these shapes is wrong, and that is the point: the same label covers markup that behaves quite differently.
Two limits, stated plainly. This is homepages only, one request each, one day — a homepage is not a site, and a site that runs connected markup on article templates may show flat markup on the front page. And this measures how common connected markup is, not what it does. We ran no test of the effect on citations, and nothing here supports or refutes the 19.72%.
Where it matters and where it does not
The case for connected markup is strongest where an entity is genuinely ambiguous to a machine. Multi-location businesses, organisations with many named practitioners, brands whose name collides with a common noun. Declaring the entities and linking them removes guesswork that a model would otherwise do by inference.
Schema App’s clearest example is not about visibility at all. A Google AI Overview stated that a Wells Fargo branch had permanently closed, working from a decades-old third-party article; after markup was deployed on the branch locator pages, the wrong result resolved and Google began citing the current pages. That is a correction of a fact, not the earning of a citation. It is the better argument of the two, and it is the one that gets less attention — particularly since a citation on its own moves less than teams expect.
Where it does not matter: a 40-page site with one location, one author and no naming collisions. There is no entity to disambiguate and no relationship a model cannot infer from the visible text. The ceiling on any gain is set by whether the pages are indexed at all. Spending a quarter on an entity graph there is spending it in the wrong place, and the documentation quoted above says so more directly than any vendor will.
The honest summary: markup earns the search features it is documented to earn. Whether it lifts AI citations is contested, the one controlled test points slightly the other way, and the vendor case studies do not settle it.
How to verify on your own site
Fetch the page and count the markup blocks:
curl -sL https://example.com/ | grep -c 'application/ld+json'
Then check whether your references resolve. Pull the JSON-LD and look for two things. First, objects that declare an identifier: an @id alongside other properties. Second, objects that are a bare reference to one: an @id and nothing else. A reference that points at an identifier you never declared is a dangling edge — it looks connected and is not.
For the rendered result, Google’s Rich Results Test reports what its parser extracted after rendering, which is what matters if your markup is injected by JavaScript. The Schema Markup Validator checks the syntax against schema.org without the search-feature filter.
If you want the same count we ran, the check is: declared identifiers, references, and the intersection of the two. The intersection is the number that separates a graph from a pile of tags.
The numbers
| Measure | Sites | Share of 20 |
|---|---|---|
| Served any JSON-LD on the homepage | 12 | 60% |
| Served none | 8 | 40% |
At least one @id reference resolving on the page | 10 | 50% |
At least one sameAs value | 10 | 50% |
| Markup present but no resolving reference | 2 | 10% |
Sample: the 20 SEO and search publications in our source list. Homepages only, one request each, 7 September 2026, curl with a desktop user agent, JSON-LD parsed from application/ld+json blocks. Largest graph observed: 28 declared entities, 19 sameAs values.
FAQ
Does adding structured data make a page eligible for AI Overviews?
No. The eligibility bar in Google’s own documentation is indexing plus the ability to show a snippet in ordinary Search, and the same page says no special markup is needed on top of that. Markup earns you the search features it is documented to earn; it does not open a separate door.
Do the studies showing no effect actually test connected markup?
The Ahrefs study tested pages that added JSON-LD and compared them against matched controls; the design does not distinguish a flat Product block from an entity graph with resolving references. That is a fair criticism of what it can tell you. It is not evidence that connected markup works — it is a reason the question is still open, which is a different claim than the one being sold.
How do I tell whether my markup is connected or flat?
Count two sets and intersect them: identifiers you declare with @id, and references to an @id elsewhere in your markup. If the intersection is zero, every object stands alone and the graph exists only in the diagram. In our sample of 20 homepages, 10 had a non-empty intersection and 2 had markup with none.
