Structured Data for AI: Definition and Stakes for AI Visibility

In the news: Since June 2026, Google and the Schema.org community publish a monthly dataset of aggregate Schema.org term usage across millions of domains, the first public view of how the web actually uses structured data (Schema.org blog, June 4, 2026).

Almost every page you will read on this topic tells you that structured data is how you get picked up by AI. That claim is now testable, and it has been tested. The result is more interesting than the promise: structured data matters, but not for the reason it is usually sold. Here is the definition, the mechanics, and the evidence.

Structured data, defined in one line

Structured data is a standardised block of code added to a web page that states, in machine-readable form, what the page is about: that this string is a price, that one an author, this date an event. Humans read the visible page. Machines read the structured data.

A web page is, to a machine, a wall of text wrapped in presentation tags. A human glancing at a product page instantly knows that "89" is a price, that the line underneath is a customer rating, that the bold string at the top is the product name. A machine has to guess. Structured data removes the guessing by restating the same facts in a fixed, agreed format.

The important word in that definition is standardised. The value is not that you describe your page, it is that you describe it in the same words as everybody else, so that any machine which has learned the vocabulary can read any site that uses it. That shared vocabulary is Schema.org, and it is the reason a recipe site in Lyon and one in Osaka can both be understood by the same parser.

Schema.org and JSON-LD: vocabulary against format

Schema.org is the vocabulary, the agreed words such as Product, Organization, author, price. JSON-LD is the format, the syntax you write those words in. You write Schema.org vocabulary in JSON-LD format, usually in a script tag in the page source.

People conflate these two constantly, and the confusion makes the whole subject harder than it is. Keep them apart and everything else follows.

Schema.org is a collaborative vocabulary: a long dictionary of types, such as Product, Recipe, Organization, Event, and of properties belonging to those types, such as price, author, startDate. It says nothing about how you write them down.

JSON-LD answers that second question. It is a W3C Recommendation, defined by the standards body as "a lightweight syntax to serialize Linked Data in JSON". Its practical virtue is that it sits in one self-contained block in your page source rather than being braided through your visible HTML, which is what made the older microdata and RDFa approaches so painful to maintain. Google recommends JSON-LD, and it is what nearly everyone now ships.

In practice a minimal block looks like this, and it lives in the head or body of the page without changing a pixel of what the reader sees.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Cicero Studio",
  "url": "https://cicero.studio"
}
</script>

That is the whole idea. A vocabulary, a format, and a block of code that says out loud what your page only implies. Which raises the question of who reads it, and what they do with it.

What structured data actually does in Google Search

In classic search, structured data makes you eligible for rich results: the stars, prices, images and breadcrumbs that make a listing bigger than a plain blue link. Eligibility is not entitlement, and the catalogue of supported types shrinks as often as it grows.

This is the part that has always been real and remains real. Mark up a recipe correctly and you become eligible for the recipe presentation. Mark up a product and the price and availability can surface directly in the results. Structured data does not improve your ranking as such, it changes what your listing can look like once you rank.

Two words in that paragraph carry the weight. Eligible means Google may show the enhanced presentation, not that it must. And the list of features Google supports is not a one-way ratchet, which is where a lot of teams got hurt over the last eighteen months.

The same pruning has been going on more quietly elsewhere. Google began deprecating a set of rarely used structured data types from January 2026, a move that was widely misread online as Google walking away from structured data altogether. Search Engine Journal traced that misreading back to its source, quoting Google's John Mueller putting it in proportion: "Understand that markup types come and go, but a precious few you should hold on to". That is the correct mental model. The vocabulary is stable; the features built on top of it are not. We broke down which types went in our note on the schema types Google retired in 2026.

Does structured data get you cited by AI?

On the evidence available in 2026, it does not appear to be a direct lever. Google states there is no special Schema.org markup needed for AI Overviews or AI Mode, and a controlled study of 1,885 pages that added schema found citation changes indistinguishable from zero. Pages with schema do get cited more, but that reflects the kind of site that maintains markup.

This is the section most pages on this keyword get wrong, so it is worth being slow and specific about what is actually known.

Start with the operator. Google's own documentation on AI features is unusually blunt: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." There is no hedging in that sentence. Whatever structured data does for you, appearing in AI Overviews is not gated behind a special markup type.

Then the measurement. In May 2026 Ahrefs published the study the industry had been arguing about without data. Louise Linehan and Xibeijia Guan tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 and compared them against 4,000 matched control pages. The results: +2.4% on Google AI Mode, +2.2% on ChatGPT, and -4.6% on AI Overviews. The first two are statistically indistinguishable from zero. The third is small, significant, and points the wrong way for anyone selling schema as an AI-visibility lever.

The interesting part is what the same team found before running the experiment. Across an initial analysis of roughly six million URLs, pages cited by AI were about three times more likely to carry JSON-LD than uncited pages. That is a large correlation, and it is exactly the statistic quoted in most articles recommending schema for AI. The experiment is what separates the two readings: sites that maintain structured data are also the sites that maintain everything else, publish authoritative content and earn links. The markup travels with the quality; it does not appear to manufacture it.

The Ahrefs report also relays testing by searchVIU on whether AI systems parse markup when they fetch a page live. The finding was that the systems tested extracted visible HTML content, with JSON-LD and hidden markup ignored. Treat that as one team's test rather than a settled law of the universe, but it is consistent with the citation numbers, and it suggests a simple discipline: if a fact matters, it belongs in the visible text, not only in your markup.

So why keep doing it?

Because structured data does three things that survive the evidence above: it earns rich results on supported types, it disambiguates your entity for knowledge graphs, and it states your facts unambiguously for any machine that reads them. None of those is glamorous. All of them compound.

Once you stop expecting markup to be a citation switch, its real value becomes easy to see.

1. Rich results, on the types that still have them

Product, Article, BreadcrumbList, Event, Recipe, Video: where Google still supports a presentation, the markup remains the ticket to it. A listing carrying a price and a rating occupies more of the screen and reads as more substantial than the plain link beside it. That is a straightforward, measurable return, and it is unaffected by anything the AI research says.

2. Entity disambiguation

This is the underrated one. Structured data is where you state that your organisation is this entity, with this identity, these profiles, this founder, not the similarly named business three regions over. Machines resolve entities by cross-referencing signals, and explicit markup is the cheapest signal you can hand them. It is the practical face of entity SEO, and it matters more the more generic your brand name is.

3. Unambiguous facts

Even where a system reads only visible text, having a canonical machine-readable statement of your prices, hours, authors and dates keeps your own house consistent. It is what feeds your knowledge panel, your voice assistant answers, and every downstream integration you have not thought of yet. And it costs almost nothing once it is generated by your CMS rather than hand-written.

Notice what all three have in common: they are about being understood, not about being promoted. That distinction is the whole subject.

The types that still earn their keep

Prioritise the types tied to a feature Google currently supports and to your actual business, and check the documentation before adding one. Marking up everything the vocabulary allows is effort spent on features that may not exist.

One rule beats every list, including this one: your structured data must match your visible page. Google's guidance on AI features closes with the reminder to ensure "your structured data matches the visible text on the page". Markup that describes a rating you do not display, or a price you do not charge, is not an optimisation. It is a discrepancy, and discrepancies are what enforcement systems are built to find.

What structured data does not do

Structured data is not a ranking factor, not a shortcut to AI citations, and not a substitute for content. It describes what is on the page. It cannot improve what is on the page.

• It does not rank you. Markup makes you eligible for a presentation once you rank. It does not move you up the page. A well-marked-up page with nothing to say stays a page with nothing to say.
• It does not manufacture authority. The three-times correlation with AI citations belongs to the sites, not to the markup. You cannot borrow the correlation by copying the tag.
• It does not protect a feature. Anything built on a specific rich result inherits that feature's lifespan, as the FAQ chip demonstrated on May 7, 2026.
• It does not replace visible text. If systems read the rendered page, a fact that exists only in your JSON-LD is a fact you have not really published.

Where Cicero Studio fits

Cicero Studio treats structured data as infrastructure: generated systematically, kept consistent with the visible page, and never sold as a growth lever. It starts with a free audit, then editorial production and automated semantic internal linking, run as one loop.

Being straight about this costs us an easy pitch. "Add schema and the AIs will cite you" is a simple thing to sell, and we could sell it. Across the 1211 SEO/GEO audits produced by Cicero Studio, the pattern I keep seeing is that markup is rarely the binding constraint. In our experience it is usually the fifth or sixth problem on a site, sitting behind pages nobody can find, thin answers to questions nobody asked, and an entity Google cannot resolve. Fixing the markup on a page that does not deserve to rank changes nothing, and I would rather tell a prospect that in the first call than in the sixth month.

So the method runs in that order: understand what is actually blocking the site, then produce content worth citing, then wire it together.

That is the promise in one line: agency-quality work, software-grade productivity. If you prefer the French-language treatment of the model, our agence GEO pillar covers it in depth, and our English breakdown of an SEO and GEO audit shows exactly what we check.

Alexis Dollé

CEO & Founder

Growth and SEO content strategist, I founded Cicéro to help businesses build lasting organic visibility, on Google and in AI-generated answers alike. Day to day, I run our clients' audits, which means I spend a lot of time telling people that the tactic they read about last week is not their problem. We put AI to work for production, never in place of expertise.

LinkedIn →

Resources to go further

We document our approach in the open, because published work with its sources beats any sales deck. The 508 articles published on cicero.studio (267 FR, 241 EN) are where we work through the answer-visibility puzzle in public. Each link below digs into one piece of it; the pillar is in French, our home market, and is flagged as such.