Structured Data for AI: Does Schema Help You Get Cited?

Structured data does not make ChatGPT, Perplexity or Google AI cite you. Google says its AI features need no special markup, and the one large before-and-after study we found saw no citation lift. Ship clean schema anyway, because it labels who you are and what you sell. Just never let a fact live only in JSON-LD.

This guide separates what is documented, what has been measured and what nobody outside the labs knows. It reflects public documentation and studies as of October 2026.

What is structured data, and why would AI care?

Structured data is machine-readable labeling added to a page, almost always as JSON-LD using the schema.org vocabulary. It tells software that "49" is a price and "AEOeye" is an organization, instead of leaving it to guess from prose.

Google says it uses markup to understand the page and to learn about the people, books and companies named in it (Google Search Central). That is entity disambiguation: is this Acme the software company or the plumber? It is the one job schema does reliably, and it is the core of entity SEO.

Google recommends JSON-LD over Microdata and RDFa because it is easiest to maintain at scale. Our glossary entry has the short definition.

Do AI assistants actually use schema markup?

Only partly, and the documented part is small. Google says its AI features need no special markup, and Microsoft recommends schema for AI search. OpenAI's and Perplexity's crawler documentation says nothing about it, and no lab documents training use. Everything else is inference or third-party testing.

Engine	What the vendor documents	Status
Google AI Overviews and AI Mode	Structured data isn't required, no special markup exists, and it must match visible text (AI features)	Documented
Microsoft Copilot and Bing	A Bing product manager lists schema among core practices for AI answers (Microsoft Advertising, October 2025)	Recommended, mechanism unstated
ChatGPT search	OpenAI's bot page covers OAI-SearchBot access, not markup	Not documented
Perplexity	Perplexity's bot guide covers robots.txt and IP ranges, not markup	Not documented
Model training	Web Data Commons extracts JSON-LD from Common Crawl, but no lab says it uses it	Unknown

We checked the OpenAI and Perplexity pages on October 3, 2026 for the words schema, JSON-LD and structured data; neither page's documentation text mentions them.

What about live fetching? On October 30, 2025, searchVIU put a price only in JSON-LD on a test page and asked five AI systems for it. None reported that price. One page and one day prove little, and searchVIU notes markup may still matter at indexing, which the test could not see.

Our reading: treat JSON-LD as an upstream hint for classification and entity matching, not a storage place for facts you want quoted. If a number matters, it must also be visible text.

Does adding schema increase AI citations?

There is no clear evidence that it does. Ahrefs tracked 1,885 pages that first added JSON-LD between August 2025 and March 2026 against 4,000 control pages. Citations moved −4.6% in Google AI Overviews (small but significant), +2.4% in AI Mode and +2.2% in ChatGPT; the last two were indistinguishable from zero.

Three caveats keep that honest:

• Every page already had 100+ AI Overview citations in February 2025, so the study says nothing about pages AI has not yet seen.
• It measured 30 days, pooled all types, and tested only JSON-LD present in the page's HTML.
• Across 6 million URLs, AI-cited pages were almost three times as likely to carry JSON-LD. Ahrefs calls that correlation, since well-run sites ship schema and also earn links.

Our position: schema is a symptom of a well-maintained site, not a lever. Spend a day on it, not a quarter.

Which schema types matter most?

Start with Organization, then the type that describes what you sell, then Article. Add BreadcrumbList cheaply and treat FAQPage as optional. This order follows what Google still documents and what defines your entity. It is not ranked by citation lift, because none has been shown.

Priority	Type	Where	Notes
1	Organization with sameAs and logo	Homepage or about page	Google recommends it there, not on every page; no required properties; use the most specific subtype
2	SoftwareApplication or Product, with Offer	The page that sells it	Google's software app rich result needs name, offers.price and a rating or review; never invent ratings
3	Article or BlogPosting	Every article	Google recommends author and dateModified
4	BreadcrumbList	Inner pages	Cheap, low risk
5	FAQPage	Only pages with visible Q&A	Google's changelog says FAQ rich results stopped on May 7, 2026

Google's rich result gallery, updated June 2026, no longer lists FAQ or HowTo.

In AEOeye's crawler access study (October 2026, 391 well-known sites, 312 homepages fetched), 73.3% of SaaS homepages (129 of 176) had Organization markup and 68.8% had sameAs. FAQPage appeared on only 7.4% (13 of 176), but on 42.9% of AI-visibility vendors (18 of 42). On e-commerce homepages, Product, SoftwareApplication or Offer markup appeared on 3.3% (2 of 61); it belongs on product pages, which we did not test.

It is a convenience sample of established brands, homepages only. Per-site data is in the study CSV, and the method is on the crawler directory hub. For the wider web, the 2024 Web Almanac found JSON-LD on 41% of pages but Organization on only 7.16% of mobile pages.

What does correct JSON-LD look like?

Correct JSON-LD sits in the server-delivered HTML, uses https://schema.org as its @context, and describes only what a visitor can see on that page. Here is what aeoeye.com ships on its homepage as of October 3, 2026, trimmed to the essentials.

Organization:

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "AEOeye",
  "url": "https://aeoeye.com",
  "logo": "https://aeoeye.com/icon.svg",
  "sameAs": [
    "https://www.pinterest.com/AEOeye01/",
    "https://www.reddit.com/user/AEOeye/"
  ]
}

SoftwareApplication with an Offer, which is true because the free audit costs nothing:

{
  "@context": "https://schema.org",
  "@type": "SoftwareApplication",
  "name": "AEOeye",
  "applicationCategory": "BusinessApplication",
  "operatingSystem": "Web",
  "url": "https://aeoeye.com",
  "offers": { "@type": "Offer", "price": "0", "priceCurrency": "USD" }
}

An Article template, with placeholder values you replace from your own page data:

{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "Your visible H1",
  "datePublished": "2026-01-15",
  "dateModified": "2026-10-03",
  "author": { "@type": "Organization", "name": "Your Brand", "url": "https://example.com" }
}

Four rules go with them:

1. Render it on the server. Vercel's December 2024 analysis found ChatGPT's and Claude's crawlers do not execute JavaScript. Google supports Tag Manager injection, but warns generated Product markup can make Shopping crawls less reliable.
2. Point sameAs at real identity pages. Schema.org defines it as a page that unambiguously indicates identity, such as Wikipedia, Wikidata or an official website. We list the two profiles we run, no more.
3. Mirror visible Q&A. Our homepage FAQPage block repeats the questions and answers printed on the page, which we checked on October 3, 2026.
4. Skip what you cannot back. No aggregateRating without real reviews.

What is schema drift, and how do you prevent it?

Schema drift is markup that stops matching the page it describes: a stale price, a deleted FAQ, a departed author. It is easy to cause, because templates and plugins emit markup that nobody rechecks when visible content changes.

Google's guidelines say not to mark up content readers cannot see, and breaking that can prevent a rich result or get markup treated as spam. A structured data manual action removes rich result eligibility; Google says it does not change web ranking. For AI, the risk is simpler: you hand machines a confident wrong fact.

To prevent it:

1. Generate markup from the same data that renders the page, never from a parallel copy.
2. Validate on every deploy and fail the build on errors.
3. Once a month, diff your top pages' JSON-LD against their visible text.
4. Delete FAQPage from pages that no longer show Q&A.
5. Review sameAs and Offer whenever pricing, branding or profiles change.

What should you fix before schema?

Fix visible, quotable text first. Microsoft's Bing guidance recommends self-contained phrasing and warns against hiding answers in tabs or expandable menus, because AI systems may not render them. Markup cannot rescue a page with nothing worth quoting.

That means headings that mirror real questions, a direct answer under each, and consistent brand names. Our guides on optimizing content for AI search and what makes content quotable by AI cover this. Google's AI guide also says Google Search does not use llms.txt; our view is in Is llms.txt worth it?.

How do you check your structured data?

Check three things: the markup is in the raw HTML, it validates, and it matches the visible page. Then measure outcomes, because valid markup says nothing about whether any engine recommends you.

1. Raw HTML. Run curl -s https://your-site.com/ | grep -c 'application/ld+json'. Zero means crawlers that skip JavaScript see nothing.
2. Validity. Use Google's Rich Results Test for Google features and the Schema Markup Validator for general syntax. Google also advises watching Search Console's rich result reports after deployment.
3. Match. Compare each price, date, author and question to the page.
4. Outcome. Ask the buyer questions your customers ask and see who gets named. Our free audit does that for three questions on ChatGPT.

To generate valid blocks, use the free AI Schema Generator.