Does schema markup make AI cite you? We measured it on 2,047 sites

In December 2025 a researcher built a test page for a made-up product — gummy bears — and hid one fact inside the code. The price of one variant, €8.99, existed only in the JSON-LD block. Nowhere visible on the page. Then he asked ChatGPT, Claude, Gemini, Perplexity and Google AI Mode what that variant cost. None of the five found it.

The experiment comes from searchVIU, and it does not prove schema markup is useless. It proves something narrower and more awkward: when a chatbot opens your page live, it reads the visible HTML, and the JSON-LD goes past it. Two much larger measurements have since landed pointing the same way. One is Ahrefs'. The other is ours, across 2,047 sites.

The causal test the industry was missing

Until May 2026 the argument about schema and AI visibility ran on anecdotes and on the intuition that a properly labelled fact must be worth more than a loose one. Ahrefs finally built a decent design around it: 1,885 pages that added JSON-LD between August 2025 and March 2026, matched against roughly 4,000 control pages with similar citation levels beforehand, measured 30 days before and 30 days after.

The results: AI Overviews −4.6%, AI Mode +2.4%, ChatGPT +2.2%. The last two are indistinguishable from zero. The first, the only one with statistical significance, points the opposite way from what gets promised in sales decks.

One line in the methodology changes how the whole thing should be read, and it is the reason we went looking at our own numbers. Every page in that dataset had more than a hundred AI Overview citations in February 2025, before anyone touched the schema. Ahrefs measured what schema does for pages AI was already citing heavily. Clean experiment, legitimate question. It also leaves out the population most business owners belong to: sites that get no citations at all.

There is a mechanical explanation sitting underneath both results, and searchVIU spelled it out. Structured data does different work at different stages. When Google and Bing crawl and index a page they parse the JSON-LD, and those indexes feed AI Overviews and the assistants built on top of them. When ChatGPT or Claude fetch a URL directly in the middle of a conversation, they take the rendered text and little else. The same markup can count on Monday and be irrelevant on Tuesday, depending on which route the answer travelled. Worth knowing before anyone budgets for markup as the thing standing between them and a citation.

Google wrote it down, with a date on it

While the industry argued, Google published its own answer, and almost nobody quoted the whole thing. Its guidance "Optimizing your website for generative AI features on Google Search," last updated July 10, 2026, is blunt on three points currently sold as services:

"Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add. However, it's a good idea to continue using it as part of your overall SEO strategy."

The same page dismisses machine-readable files — "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search" — and dismisses chunking too: there is no requirement to break content into tiny pieces, because its systems can handle several topics on one page and surface the relevant part.

The one thing Google does ask, in its documentation on AI features, is that structured data match the visible text on the page. Which is the same condition the searchVIU test forces from the other direction: if a fact lives only in the JSON-LD, there are systems that will never see it.

The 2,047 sites Ahrefs did not measure

Our sample is a different animal, which is what makes it useful alongside theirs. Between May 14 and July 28, 2026 we ran 2,324 audits across 2,021 unique domains. Each one fires real questions at a model with live web search and stores the sources that model actually opened. Most of these are ordinary businesses — clinics, shops, agencies, workshops. Across the whole set, 76.6% get no mention at all when you ask about their category.

Here is JSON-LD presence crossed against mention rate on the 2,047 sites where we had both:

Base: the 2,047 sites where both data points were available, out of 2,324 Inspeccia audits (May 14 – July 28, 2026). "Invisible" means the site got no mention at all when the model was asked about its category.

The gap is real: 5.7 percentage points. Not noise. But before anyone turns it into a slide, it is worth saying plainly what it is and what it is not.

It is a correlation, and we cannot untangle it from anything. Sites that bother implementing structured data also tend to be bigger, older, better funded, staffed with someone technical, publishing for years and already mentioned in the press. Any one of those explains 5.7 points as well as schema does. We have no way to separate the effect of JSON-LD from the effect of simply being a better-maintained site, and anyone claiming they can do that with observational data is selling something. Ahrefs intervened and used matched controls, so their null result carries more weight than our positive gap.

The limits of our measurement, so you can discount them: a single LLM provider, a cap of twelve source domains stored per analysis, and self-declared categories.

The honest headline is the other column. Three out of four sites with correct structured data are still invisible to AI. If your schema is in place and you do not show up, schema was not the missing piece.

We watched this same film with llms.txt

A year ago the promise was a text file at the root of your domain. Hundreds of guides got written, audits got sold, several CMS platforms shipped the button. In June 2026 Ahrefs measured what had actually happened: across 137,210 domains, 28% published a valid llms.txt, and 97% of those files received not a single request in May 2026. Of the requests that did arrive, 96% came from bots, and the largest identified category — 21.7% — was SEO audit tools checking whether the file existed. An industry studying itself.

Google's John Mueller summed it up on June 2, 2026: the whole thing is "purely speculative for now," and the file has existed for years without any AI system using it. The full account is in our llms.txt guide.

The pattern repeats often enough to deserve a name. A plausible technical format appears, gets assigned a function nobody documented, sells for eighteen months, and then somebody measures and the effect is zero. Sitemaps, robots.txt and canonical tags were new formats once too, but they arrived with documentation from the engine consuming them and a function you could verify. That is the bar.

What schema markup is actually for

None of this argues for ripping it out. Schema markup does three documented things, and none of them is "make ChatGPT cite you."

It qualifies you for Google rich results. Star ratings, prices, availability, breadcrumbs, events, recipes, FAQs where they still render. This is a formatting requirement rather than a quality signal: without correct markup there is no rich result, with it there is eligibility. Reason enough to implement it properly.

It holds your entity together. An Organization block with sameAs pointing at your official profiles gives any system a cheap way to confirm that the company on the site, the one on LinkedIn and the one in the listing are the same company. Across our audits we logged 50 cases of entity confusion, where the model blends a business with a similarly named one, and consistent declared identity has something real to contribute there.

It forces your data to match the visible text. This is the underrated part and the only thing Google asks for explicitly. When you mark up hours, prices or an address and those same facts sit in the HTML, you close the gap the searchVIU test exposed. If you need the mechanics — where each block goes, which fields cannot be missing — they are in where to add your business JSON-LD and in our guide to schema types for AI search, which still stands as an implementation manual. This article is the measurement that puts a ceiling on what to expect from it.

Where the lever actually is

If schema does not explain why you are absent, something has to. Our data points somewhere specific: AI cites an average of 2.65 sources per brand, and in 26.9% of cases it does not open the company's own website at all. In more than one answer in four, everything the model says about a business comes from reviews, directories, local press, forums and scam-check sites.

That is where the argument closes itself. If the model never reads your site, the JSON-LD on your site cannot help you, however immaculate it is. The useful question stops being which markup you are missing and becomes which pages the model is reading when someone asks about your category, and whether you are on them. We measured and broke that down in what sources AI actually checks when someone asks about your business.

The structural work on your own site still matters, for reasons that predate AI: be crawlable, load, put the content in the HTML rather than only in JavaScript, make each page state clearly what it covers. That is in our technical AISO guide. None of it is glamorous and all of it can be verified, which is more than can be said for the markup that was going to change everything.

One caveat covering every figure here, ours included: they are a snapshot with a date on them. Platforms shift their preferred sources from one quarter to the next, and any citation percentage ages fast. What ages more slowly is the method — ask for the measurement before you pay for the implementation.

Frequently asked questions