Does schema markup get you cited by AI? We checked the data

A whole cottage industry now sells schema markup as the way to get cited by ChatGPT and Google's AI answers. The claim is tidy, technical and satisfying. It is also, on the current evidence, mostly wrong. A controlled study of 1,885 pages found almost no citation lift, Google says no special schema is required, and one experiment found most AI engines cannot even read your markup on demand. Here is what the data actually shows, and what schema is genuinely good for.

The claim, and why it is so appealing

Open any recent guide to answer engine optimisation and you will meet the same promise: add the right schema markup and the AI engines will start citing you. It is an appealing story because it turns a hard, fuzzy problem, being named in an AI answer, into a concrete technical checklist. Add FAQPage here, Article there, mark up your organisation, and visibility follows. Some guides even attach precise-sounding numbers, claiming pages with structured data are dozens of percent more likely to be cited.

We went looking for the evidence behind those numbers, because Stellarcast's whole job is to be right about how AI answers work, not to repeat the loudest claim. What we found is that the strongest available data points the other way. Schema markup correlates with being cited, but the best controlled test to date shows it does not appear to cause the citation. The tidy checklist is selling a lever that, on the evidence, is not connected to the thing it claims to move.

What the controlled study found

The most rigorous test we found is from Ahrefs, published in 2026. Rather than eyeballing which cited pages happen to have schema, they ran a proper before-and-after comparison. They tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, and compared each against matched control pages, roughly three controls per treated URL, that never added schema and had similar citation levels beforehand. That control group is the important part: it isolates the effect of the schema itself from everything else a site might be doing.

The result was close to nothing. Citations from Google AI Overviews moved by about minus 4.6 percent, a small but real decline. AI Mode moved about plus 2.4 percent and ChatGPT about plus 2.2 percent, both small enough to be indistinguishable from noise. In plain terms, adding schema did not produce the citation gains the marketing promises. If anything, on the one surface where the change was statistically clean, it edged down.

"Adding schema to 1,885 pages moved AI citations by a rounding error. The lever most guides sell you turned out not to be connected to the machine."

Want to know whether AI engines cite your pages today, with or without schema? Get a free visibility audit and see exactly how ChatGPT, Gemini and Perplexity describe you.

The correlation trap

Here is where the myth gets its fuel. In that same Ahrefs data, about 53 percent of AI-cited pages carried schema, roughly three times the rate of non-cited pages. Read carelessly, that looks like a smoking gun: cited pages have schema, therefore schema causes citations. It is exactly the kind of statistic that gets screenshotted into a slide.

But the before-and-after test already ruled that reading out. The reason cited pages tend to have schema is that the sites which bother to add schema are usually the sites doing everything else well too: they publish authoritative content, keep it fresh, structure it clearly, and earn links and mentions. Schema is a marker of a well-maintained page, not the cause of the citation. Confusing the two is the single most common mistake in AEO advice, and it is why a real correlation gets sold as a false promise.

Google says no special schema is required

You do not have to take a third-party study's word for it, because Google has stated the position directly. In its own documentation on AI features and your website, Google writes that you do not need to create new machine-readable files or markup to appear in AI Overviews and AI Mode, and that there is no special schema.org structured data that you need to add. The stated requirement is simpler: a page has to be indexed and eligible to be shown in Google Search with a snippet.

That is about as clear as platform guidance gets. The engine that runs the largest AI answer surface in the world is telling you, in writing, that no schema unlocks it. The bar is being crawlable, indexed and snippet-eligible, which is a content and structure problem, not a markup one.

Most engines cannot even read your markup on demand

There is a second, more mechanical problem with the schema-earns-citations story. In an experiment run between December 2025 and March 2026, the monitoring tool Otterly.ai tested whether seven major platforms could actually fetch and interpret a page's schema when asked directly. The finding was blunt: six of the seven, ChatGPT, Google AI Overviews, AI Mode, Perplexity, Copilot and Claude, could not reliably fetch or correctly interpret the JSON-LD. Only Gemini retrieved it, and in one case AI Mode described schema that did not exist on the page at all.

This is a single-vendor, single-site test, so treat it as directional rather than the last word. But it undercuts the mechanism the myth depends on. If the engine never reliably reads your raw markup in the first place, it is hard for that markup to be what wins you the citation. It fits the controlled-study result: the effect is small because the causal path is weak.

What schema is actually good for

None of this means you should rip your schema out. It means you should be honest about what it does. Schema is a clean, standard, machine-readable way to state facts about your page and your organisation, and that has real value even if it is not a citation lever.

• Understanding and entity disambiguation. Microsoft has said on the record, at SMX Munich in 2025, that schema markup helps its LLMs understand content. Understanding is not citation, but being correctly understood is the floor you build everything else on.
• Stating facts unambiguously. Marking up your organisation, authorship, dates and products gives machines an unambiguous version of details they would otherwise have to infer from prose. That reduces the chance an engine gets your basic facts wrong.
• It does not hurt. Google is clear that unused structured data on a page is not a negative signal. Valid schema is low-cost hygiene, so keep it clean and accurate.
• Traditional rich results, where they still exist. Some schema types still drive visual search features. Just note that two popular ones no longer do: HowTo rich results were retired in 2023, and FAQ rich results stopped showing in May 2026.

Where the real citation leverage is

If schema is not the lever, what is? The same things the controlled study kept pointing back to as the likely real drivers. These are less tidy than pasting JSON-LD, which is precisely why they work.

• Be the clearest source about your own facts. State what you do, who it is for and why you are credible in plain, quotable sentences an engine can lift without interpretation.
• Earn corroboration. Engines trust a claim more when independent, reputable sources repeat it. Being described consistently across the places your buyers already trust does more than any markup.
• Structure for retrieval. Answer-first sections, question-shaped headings and self-contained passages help the retrieval stage surface your content. This is about the visible words on the page, not hidden markup.
• Keep it fresh and authoritative. The well-maintained sites that correlate with citations got there through content quality and authority, not through a schema checkbox.

A note on the numbers you will see online

As you read around this topic, you will meet confident statistics: pages with structured data are 28 to 40 percent more likely to be cited, multimedia pages 156 percent more likely, and so on. We went looking for the studies behind those figures and could not find them. They trace to marketing blogs citing each other, with no dataset, methodology or primary source underneath. On a topic where a real controlled study exists and shows almost no effect, we would treat any unsourced double-digit lift claim as marketing, not evidence. When someone cites a number, ask for the study. If there is not one, it is a slogan.

The takeaway

Schema markup is useful hygiene that helps machines understand and correctly state your facts. It is not, on the best available evidence, the thing that earns you an AI citation. A controlled 1,885-page study found almost no lift, Google says no special schema is required, and most engines cannot reliably read your markup on demand anyway. So keep your schema clean, but spend your real effort where the leverage actually is: being the clearest, best-corroborated, most quotable source about your own brand. That is what gets you named in the answer.