Schema markup for AI search: what it does, what it doesn't, and what to add

What is schema markup for AI search?

Schema markup is code that describes the things on a page in a shared vocabulary, so a machine reads "this is an Organization called Acme with this logo and these profiles" instead of guessing from the text. For AI search, people use the term for the same markup they already add for Google, applied with the hope that ChatGPT, Gemini, Perplexity and Google's AI features will understand, trust and cite the page more often.

The vocabulary comes from Schema.org, a project founded by Google, Microsoft, Yahoo and Yandex and run since April 2015 through a W3C community group. You can write it three ways: JSON-LD, Microdata or RDFa. Google's introduction to structured data says all three are "equally fine" when valid, and that Google recommends JSON-LD because it is the easiest to add and maintain at scale. JSON-LD sits in a <script type="application/ld+json"> block, separate from the visible text. That separation matters later in this guide, because some AI systems strip script blocks before a model ever sees the page.

Here is what a minimal block for a company looks like. It is the kind of markup that sits behind a knowledge graph entry.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Acme Payroll",
  "url": "https://www.acmepayroll.example",
  "logo": "https://www.acmepayroll.example/logo.png",
  "sameAs": [
    "https://www.linkedin.com/company/acmepayroll-example",
    "https://www.crunchbase.com/organization/acmepayroll-example"
  ]
}
</script>

Two terms get mixed up in most guides. Structured data is the general idea of describing content in a fixed format. Schema markup is structured data written in the Schema.org vocabulary. Every argument below is about the second one.

What do Google, Microsoft and OpenAI say about schema and AI answers?

Google says structured data is useful but not required for AI features. Microsoft says schema helps its LLMs understand content. OpenAI publishes no schema guidance for ChatGPT's organic answers and uses a separate product feed for shopping. Those three positions differ more than most "schema for AI" articles admit.

Google's page on AI features and your website is the clearest statement from any engine owner. It says there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary," and that there's "no special schema.org structured data that you need to add." The same page asks site owners to make sure structured data matches the visible text. Google's May 2025 post, Top ways to ensure your content performs well in Google's AI experiences on Search, repeats the point: structured data shares information "in a machine-readable way that our systems consider" and makes pages eligible for rich results, so follow the guidelines and keep it consistent with the page.

Microsoft has gone further. At SMX Munich in March 2025, Bing's Fabrice Canel said schema markup helps Microsoft's LLMs understand content, as Barry Schwartz reported for Search Engine Land. Canel confirmed the remark in the LinkedIn comments under David Mihm's post, and added that generative AI values fresh content, pointing site owners to IndexNow. Microsoft Advertising's October 2025 post, Optimizing Your Content for Inclusion in AI Search Answers by Krishna Madhavan, describes schema as code that "helps search engines and AI systems understand your content" and, in the same piece, warns against hiding key answers in tabs or expandable menus because AI systems may not render them.

OpenAI is the quiet one. Its crawler documentation covers access, not markup, and its commerce documentation asks merchants for a structured product feed delivered as a file, with field-by-field validation rules. In other words, when OpenAI wants structured product data, it asks for a feed instead of reading Product markup off your pages. A thread on the OpenAI developer forum asking whether ChatGPT's browsing tool extracts JSON-LD has no official answer.

Where does schema get read in the AI search pipeline?

Schema can matter at three different moments: when a search engine crawls and indexes the page, when a model's training data is built, and when an AI assistant fetches the page live to answer a question. The evidence is strongest for the first and weakest for the third, and most disagreements online come from people talking about different moments.

The SEO consultant Suganthan Mohanadasan calls these the three lives of schema markup, and the framing is useful because each stage has a different reader.

• Index time. Google and Bing parse JSON-LD when they crawl. Google uses it for rich results and to disambiguate an organization from others with similar names. AI Overviews and AI Mode draw on the same index, so anything schema adds to Google's understanding is available to them indirectly.
• Training time. Whether raw JSON-LD survives into the text a model is trained on depends on how each lab cleans its crawl, and none of the major labs documents this. Treat any claim either way as a guess.
• Retrieval time. When ChatGPT, Claude or Perplexity fetch a page to answer a question, the page usually passes through an HTML-to-text step. Script blocks, which is where JSON-LD lives, are often dropped there.

The retrieval stage has been tested. Otterly ran a three-month experiment from December 7, 2025 to March 7, 2026, and in its first test simply asked seven AI platforms to fetch a blog post and return its schema. Ahrefs, in its schema study, also describes a searchVIU experiment in which ChatGPT, Claude, Perplexity, Gemini and Google AI Mode extracted only visible HTML during direct retrieval and ignored JSON-LD, hidden Microdata and hidden RDFa.

The practical reading: if a fact only exists in your JSON-LD, most assistants reading your page live will never see it. Copilot's result is the interesting one, because Microsoft is the company saying schema helps. Both can be true. Bing can use schema when it indexes the page, while the Copilot chat session that fetches it later sees only the text.

Does schema markup increase AI citations?

Not measurably, on the best evidence available. The only large controlled test, by Ahrefs, found no lift in ChatGPT or Google AI Mode citations after pages added JSON-LD, and a small decline in AI Overviews that it could not pin on schema. Correlation studies that show cited pages carry more schema don't separate schema from everything else those sites do well.

Ahrefs started with a correlation: across 6 million URLs, AI-cited pages were almost three times more likely to have JSON-LD than pages that weren't cited, and 53% of AI-cited pages had schema. Then it ran the test that matters. Its team found 1,885 pages that added JSON-LD between August 2025 and March 2026, matched each with three control pages from other domains that had similar citation levels and never added schema, and compared citations in the 30 days before and after. The difference-in-differences result is in Figure 1.

The raw numbers show why controls matter. AI Mode citations on the treated pages grew 43% after they added schema, which would have made a great LinkedIn slide. The control pages grew almost as much, because AI Mode was expanding for everyone at the time. Once the platform trend is removed, the +43% shrinks to +2.4%.

Ahrefs lists its own limits, and they're worth reading before you dismiss schema. Every page in the study already had 100 or more AI Overview citations in February 2025, so it says nothing about pages AI systems haven't found yet. It pooled every schema type together, so a single type could still help. The window was 30 days, and it only measured JSON-LD in the HTML, not markup injected with JavaScript.

Two other studies point the same way. Search Atlas published The Limits of Schema Markup for AI Search on December 14, 2025, comparing domain-level schema coverage with visibility in OpenAI, Gemini and Perplexity answers, and found that higher schema coverage "does not reliably lead to higher visibility scores." Otterly's single-site experiment reported large gains in Google SERP features and AI Overviews after three months of schema work on its own blog, and limited impact on the other AI platforms. Both are vendors with tools to sell, and Otterly's result comes from one website, so weigh them below the Ahrefs test.

What we found on 458 pages Google's AI Overviews cited

We checked the HTML of every page Google's AI Overviews cited in our B2B citation study and recorded what structured data each one carried. Three in four cited pages had JSON-LD, but the cited pages were not more likely to carry it than the top-ranking pages Google left out, and the forum threads Google cited almost never had any.

The method, so you can judge it. The study covers 742 citations from 122 AI Overviews for B2B software queries (CRM, HR and payroll, accounting, cybersecurity and seven more categories), captured on October 9, 2026. That gives 711 unique cited URLs. We set aside the 146 YouTube URLs, which all share one page template, leaving 565. On October 20, 2026 we requested each one with a desktop Chrome user agent and parsed the raw HTML for JSON-LD blocks and Microdata. 458 pages returned the page; most of the rest (95) answered with a 403 bot block. We read raw HTML only, the same way Ahrefs did, so schema added purely by JavaScript after load is not counted. The per-URL results are in a CSV you can download.

74.0% of cited pages (339 of 458) carried at least one JSON-LD block, 8 carried Microdata only, and 111 (24.2%) had no schema we could find. Nine pages had at least one JSON-LD block that would not parse as valid JSON, and Google cited them anyway.

The split by source type is where the story changes. Publishers, agencies and software vendors almost all use schema, mostly because their CMS or SEO plugin adds it by default. Reddit threads, forum posts and social pages almost never do, and they still earned 72 cited pages in the sample.

A 74% schema rate among cited pages tells you nothing on its own, because you'd need to know the rate among pages that weren't cited. Our study also saved the organic top 10 for 40 of the queries, so we compared like with like: pages that ranked in the top 10 and got cited in that query's AI Overview against pages that ranked in the top 10 and didn't.

The cited group had slightly less schema on every row. The gap is not statistically meaningful at these sample sizes (a two-proportion test on the JSON-LD row gives z of about -1.3), so the honest conclusion is that schema didn't separate the pages Google chose from the ones it skipped. FAQPage, the type most often sold as an AI citation tactic, appeared on 20.0% of cited pages and 21.6% of skipped ones.

Which types do cited pages carry? Mostly whatever the CMS adds by default. Among the 458 cited pages, the most common were Organization (62.9%), a WebPage type (55.9%), Article or BlogPosting (52.8%), ImageObject (49.3%), Person (47.6%) and BreadcrumbList (43.9%). FAQPage appeared on 29.0%, SoftwareApplication on 7.2%, ItemList on 6.6%, reviews or ratings on 6.6%, Product on 1.7% and HowTo on 0.7%. That profile looks like Yoast, Rank Math and Webflow defaults, not like a deliberate AI strategy.

Which schema types are worth adding for AI search?

Add Organization on the homepage, Article or BlogPosting with a real author on editorial pages, BreadcrumbList site-wide, and Product or SoftwareApplication with Offer on pages that sell something. FAQPage is optional now that Google has dropped its rich result. Skip types Google has retired, and don't add markup for things the page doesn't say.

The reasoning behind each pick is about what the markup does in Google and Bing today, because that's where the evidence is.

Three Google changes are behind the last rows. In August 2023 Google limited FAQ rich results to "well-known, authoritative government and health websites" and removed How-to rich results entirely by September 13, 2023. Then its Search Central changelog announced on May 8, 2026 that the FAQ rich result would no longer appear in Google Search starting May 7, 2026, and the FAQ documentation was removed in June 2026. In June 2025 it phased out seven more types in Search results: Book Actions, Course Info, Claim Review, Estimated Salary, Learning Video, Special Announcement and Vehicle Listing. Google said at the time that the change doesn't affect ranking and that use of these types outside Google Search isn't affected.

FAQPage deserves a separate word because it's the type most "AI schema" guides push. The argument is that AI answers are question-shaped, so question-shaped markup must help. Our data doesn't support a citation effect: 20.0% of cited top-10 pages had FAQPage against 21.6% of skipped ones, and since May 2026 the markup no longer earns a rich result in Google either. A visible FAQ section is still worth having, because the question-and-answer text itself is what an assistant can lift. If your CMS already outputs FAQPage for that section, there's no need to rip it out. If you'd be inventing questions to fill markup, don't. Our glossary entry on FAQ schema covers the mechanics.

For ecommerce, the more important structured data for AI shopping is the feed, not the page. OpenAI's product feed spec sets field-level rules for products to appear in ChatGPT, and Google's own docs on generating structured data with JavaScript warn that dynamically generated Product markup makes Shopping crawls less frequent and less reliable. Keep on-page Product markup server-rendered, and treat the merchant feeds as the main channel.

How to add schema so machines can actually read it

Put JSON-LD in the server-rendered HTML, make every value match text a visitor can see, write the important facts in the visible copy as well, and validate before you ship. Those four rules cover nearly every failure we see, and they help with Google and Bing whether or not any assistant reads the markup.

1. Render it on the server. The Ahrefs study and our crawl both read raw HTML. A fetcher that doesn't run JavaScript won't see markup injected by Google Tag Manager or a client-side script. Google can render JavaScript, but its own docs flag reliability problems for Product markup added that way. View the page source (not the inspector) and search for ld+json.
2. Match the visible page. Google's structured data guidelines say not to mark up content that is hidden from users, and the AI features page asks for the same match. A price in Offer that differs from the price on the page is the most common mismatch on SaaS sites.
3. Write facts in text too. If your free trial length, founding year, certifications or integration count only appear in JSON-LD, assume ChatGPT, Claude, Perplexity and Copilot won't see them when they fetch the page. Microsoft's own advice is to keep key answers out of tabs and accordions for the same reason.
4. Use one connected graph. Give the Organization an @id and point Article publisher and author fields at it, so the page describes one company and one author instead of three disconnected objects.
5. Validate. Run the page through Google's Rich Results Test or the Schema.org validator. In our crawl, 9 of 458 cited pages had at least one JSON-LD block that wasn't valid JSON, which usually means a CMS template broke on a quote or a trailing comma.

Here is a pattern for an editorial page that ties the article, author and company together. Swap in your own values, and only include fields the page shows.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Organization",
      "@id": "https://www.acmepayroll.example/#org",
      "name": "Acme Payroll",
      "url": "https://www.acmepayroll.example/",
      "logo": "https://www.acmepayroll.example/logo.png",
      "sameAs": ["https://www.linkedin.com/company/acmepayroll-example"]
    },
    {
      "@type": "Person",
      "@id": "https://www.acmepayroll.example/team/dana-lee#person",
      "name": "Dana Lee",
      "jobTitle": "Head of Payroll Compliance",
      "url": "https://www.acmepayroll.example/team/dana-lee"
    },
    {
      "@type": "Article",
      "headline": "How to run payroll for remote employees in three states",
      "datePublished": "2026-03-02",
      "dateModified": "2026-09-14",
      "author": {"@id": "https://www.acmepayroll.example/team/dana-lee#person"},
      "publisher": {"@id": "https://www.acmepayroll.example/#org"},
      "mainEntityOfPage": "https://www.acmepayroll.example/blog/remote-payroll-three-states"
    }
  ]
}
</script>

If you don't want to write it by hand, our free schema generator builds Organization, Article, FAQPage and similar blocks in the browser. AI chatbots can also draft markup, which is one of the questions people search for, but check every value against the page: a model will happily add a rating, a price or a sameAs profile that doesn't exist.

Schema mistakes that waste the effort

Most schema problems on B2B sites aren't missing types. They're markup that says something the page doesn't, markup nobody can read, or markup added in place of work on the page itself.

• FAQ markup with no visible FAQ. It breaks Google's guidelines, and an assistant fetching the page never sees the questions. Put the questions on the page or drop the markup.
• Facts that live only in the markup. A "foundingDate", an "award" or a review count inside JSON-LD, with nothing in the text. Write it into the about page copy.
• Plugin duplicates. An SEO plugin and a theme both output Organization or Article, with different names or logos. Pick one source.
• Stale Offer prices. Pricing pages change and the hard-coded JSON-LD doesn't. If prices change often, generate the markup from the same data the page uses.
• Retired types. FAQPage, HowTo, Course Info and the other types whose rich results Google retired still validate on Schema.org, so validators won't warn you. They just won't produce a rich result in Google Search.
• Treating schema as the AI project. Teams spend a sprint on markup and none on the pages AI engines would quote. Our guide to how AI engines choose sources covers what moves those decisions.

How to test whether schema helps your own pages

Run a small matched test: add schema to a handful of pages, leave similar pages alone, and compare how both groups' AI citations change over at least 30 days. The control group is the part most teams skip, and without it you'll mistake a platform-wide swing for a schema effect.

This is a scaled-down version of the Ahrefs method, which the study itself suggests trying.

1. Pick 5 to 10 test pages that already get some AI citations, plus 5 to 10 control pages with similar citation levels.
2. Record a baseline for both groups across the AI surfaces you track, using the same prompt set. Our guide to measuring AI visibility covers building that set.
3. Add schema to the test pages only, note the date, and change nothing else on those pages during the window.
4. After 30 days (longer if you can), compare the change in each group, not the raw before and after.

The arithmetic, with invented numbers to show the method (this is a hypothetical, not a result). Say your 8 test pages were cited 120 times in the 30 days before you added Article and Organization markup, and 150 times in the 30 days after: up 25%. Your 8 control pages went from 110 to 135 citations in the same windows: up 22.7%. The schema effect is the difference between the two changes, about 2.3 points, which on samples this small is noise. Without the control group you'd have reported a 25% lift.

If your treated pages clearly beat the controls on more than one AI surface, and the result holds when you repeat it, you've found something the published studies haven't. If they don't, you've saved yourself a site-wide markup project.

Frequently asked questions