Structured Data for AI Search

Schema markup gives machines a labeled version of what a page already says. It will not, by itself, earn you more citations in ChatGPT or Google AI Overviews. The strongest test run so far tracked 1,885 pages that added JSON-LD and found no meaningful citation gain on any platform. Write schema anyway, for entity clarity and for rich results, and because Bing says its own models read it. Just stop selling it internally as an AI visibility lever.

What Structured Data Actually Is

Structured data is a labeled description of a page, written for software rather than for readers. You state the facts in a fixed vocabulary: this is an Article, the author is a Person named Josh Dennis, the publisher is an Organization called Emerald Beacon, it went live on this date.

The vocabulary comes from schema.org, the shared standard Google, Microsoft, Yahoo and Yandex launched together in 2011. The format to use is JSON-LD, a block of JSON inside a script tag. Microdata and RDFa still exist. Ignore them. Google names JSON-LD as its recommended format, and it carries one practical advantage that outweighs any theory: the markup lives in a single block you can read, diff and validate, rather than scattered across your HTML attributes.

That is the whole idea. No magic in it.

What the Evidence Actually Shows

Most advice on this topic skips the part where someone checked, so here are the numbers.

The correlation looks fantastic. Ahrefs scanned six million URLs and found that pages cited in AI answers were almost three times more likely to carry JSON-LD than pages nobody cited. Among AI-cited pages, 53% had schema. Stop reading there and the conclusion writes itself.

Then Ahrefs tested the claim properly. In a study published on May 11, 2026, the team took 1,885 pages that added JSON-LD between August 2025 and March 2026 and matched them against roughly 4,000 control pages that never added any. Every page in the set already carried 100-plus AI Overview citations. They measured a 30-day window on either side of each change, then ran a difference-in-differences analysis to strip out platform-wide trends.

The result: nothing. Google AI Mode moved 2.4% up and ChatGPT 2.2% up, both statistically indistinguishable from zero. Google AI Overviews went the wrong way, down 4.6%, a drop big enough to clear significance, though the authors carefully decline to pin it on schema. Raw AI Mode growth looked like 43% until they removed the trend that had lifted everybody.

Their own summary is blunt. If the goal is more AI citations for pages the models already see, “our data doesn't support that bet.”

Other work points the other way, and all of it is correlational. AirOps, in a report with Kevin Indig covering 16,851 ChatGPT queries across 353,799 pages, found pages with JSON-LD cited 38.5% of the time against 32.0% without: a 6.5 point gap. A paper on arXiv called GEO-16 audited 1,100 URLs and 1,702 citations across Brave, Google AI Overviews and Perplexity, and ranked structured data its third strongest pillar. Both deserve a read. Neither separates cause from quality, which is the exact gap the Ahrefs experiment set out to close. Well-run sites tend to carry schema and tend to get cited. The markup may be a symptom rather than a cause.

Google's own documentation says the quiet part out loud. On eligibility for AI Overviews and AI Mode: “You don't need to create new machine readable files, AI text files, or markup to appear in these features.” And then, plainer still: “There's also no special schema.org structured data that you need to add.”

Microsoft reads it differently. Fabrice Canel confirmed on stage at SMX Munich in March 2025 that Bing feeds schema markup to its own language models to help them understand content. He named no types. Microsoft has published nothing since that narrows it down. OpenAI, Anthropic and Perplexity have said nothing either way.

So one platform says it reads your markup, one says you do not need any, and the only controlled experiment found no lift. That is the honest state of play, and it should change what you promise your CEO.

What Schema Is Still For

Three jobs, all of them real.

• Rich results. Reviews, products, recipes, events, job postings, breadcrumbs. These are the visible search features that require markup, and they still move clicks.
• Entity clarity. This one matters for AI search, and it is not the same job as citation. A model has to resolve that Emerald Beacon, the fractional CMO firm, is one specific thing with one website, one set of services and one founder. Markup is the cheapest way to assert those facts in machine-readable form and tie them to the places that corroborate them, using sameAs.
• Not contradicting yourself. Google's AI features page lists exactly one structured data best practice: “Making sure your structured data matches the visible text on the page.” Markup that disagrees with the page is worse than no markup at all.

The Types Worth Writing

Most sites need five. Not fifty.

1. Organization, once, sitewide. Give it name, url, logo, description, a real address if you have one, and sameAs links pointing at LinkedIn, Crunchbase and any profile a model might use to confirm you exist. Assign it a stable identifier such as https://example.com/#organization.
2. WebSite, once, sitewide, with its publisher pointing at that Organization identifier.
3. Article or BlogPosting, on every post: headline, description, datePublished, dateModified, image, and an author that is a Person object rather than a bare string.
4. BreadcrumbList, on every page below the homepage. It earns the visible breadcrumb in Google results and tells a crawler where the page sits.
5. Service or Product, on each commercial page, describing what you sell and who provides it.

Then stop. A sixth type adds work, not visibility.

The FAQPage Exception

FAQPage deserves its own note, because the reason to use it changed. Google restricted FAQ rich results to government and health sites back in 2023, then ended them for everyone on May 7, 2026. Search Console reporting and Rich Results Test support come out in June 2026, and the Search Console API drops it in August. Google says existing markup can stay, causes no harm, and produces nothing visible. So keep FAQPage if it labels a genuine question block your readers can see. Delete it if you added it hunting a search feature that no longer exists.

How to Implement It

1. Write one entity graph, not seven loose blocks. Use a graph node and give every entity an identifier. The Article points to the WebPage, the WebPage to the WebSite, and everything to a single Organization. A model following those pointers can tell that twenty-eight blog posts share one publisher.
2. Template it. Put the Organization and WebSite nodes in the shared head of every page, then generate the per-page nodes from the page's own fields. Hand-written schema drifts inside a quarter.
3. Only state what the page shows. If the markup claims a 4.8 rating, a reader must be able to find that rating on the page. This rule has teeth. Google issues manual actions against structured data spam, and the markup simply stops working.
4. Keep dates honest. The dateModified field should change when the content changes, not on every build. A site where every page claims an update this morning teaches crawlers to distrust the field entirely.
5. Make the author a real entity. An author as a bare string gives a model nothing to resolve. An author as a Person with a url and sameAs links gives it something to verify.
6. Mirror your markup in prose. A model reading the page needs the claim in the body text as well. Schema labels what is there. It does not substitute for it.
7. Validate before you ship. Google's Rich Results Test reports which features a page qualifies for. The Schema.org validator catches vocabulary errors Google's tool waves through. Run both, every template, once.

What to Check After Launch

Search Console's rich result reports show errors and warnings by type, and that is where a broken template surfaces first. Pull a crawl monthly and confirm every post still carries an Article node with the right publisher identifier. Then watch the thing you actually care about. It is not markup coverage. It is whether assistants name you, and how often, and in answer to which questions. We cover that measurement layer in Generative Engine Optimization: The 2026 Playbook.

If citations are the goal, spend the day you saved on the levers with evidence behind them: answer-first passages a model can quote cleanly, third-party coverage that corroborates your claims, and a crawler nobody blocked at the CDN. How to Get Cited by ChatGPT, Perplexity and AI Overviews walks through that sequence in order. The same skepticism applies to the file everyone added last year, which we measured in llms.txt: What It Is and Why Your Site Needs One.

The Bottom Line

Write schema. Five types, one graph, generated from real page fields, matching what readers see. It costs a day of engineering time and it earns that back in rich results and entity clarity.

Just do not put it on a board slide as the reason AI will start recommending you. The one controlled experiment anybody has run says it will not. Our AI SEO practice treats markup as hygiene, the floor rather than the strategy, and the strategy itself sits in the GEO framework: what models can understand, what they can verify, and who else vouches for you.