Does schema markup help you get cited by AI assistants?

Schema markup is often sold as a quick win for AI search. Whether it helps you get cited depends on how an assistant reaches your page, and the evidence differs a great deal from one route to the next. This post checks it route by route, then sets out where markup plainly earns its place. The basics of entities, JSON-LD and rich results are in the entities and structured data guide.

The first route is indirect: Google and Bing parse markup when they index a page, and their indexes feed AI Overviews, AI Mode, Copilot and other products that ground answers in search results. The second is a live fetch, when an assistant opens your page during a conversation and converts the HTML into text for the model. The third is training, long before anyone asks a question. The AI search guide explains how training and live retrieval differ.

This is the route where markup is plainly read. Google says its generative AI features are rooted in its core ranking and quality systems and retrieve pages from its Search index, and that the way Search finds and processes your pages remains the core of how its AI systems access them.¹ Search uses the structured data it finds to understand a page and to gather information about the web and the world in general.² So markup can shape how Google understands a page, and the entity behind it, before any AI feature uses it.

Google is just as clear that none of this is required. Structured data is not needed for its generative AI features, there is no special schema.org markup to add, and it is still worth using as part of ordinary SEO.¹ Its page on AI features asks only that structured data match the visible text on the page.³

Bing goes further. Its webmaster guidelines say Bing and Copilot rely on the same crawling, indexing and ranking foundation as traditional search, and that structured data may support clearer grounding but does not guarantee visibility or grounding traffic. Markup must reflect visible content, and misleading markup may be ignored. The guidelines cover Bing search, Copilot and grounding API results alike.⁴ That is the strongest platform support for markup in AI answers, and it is carefully hedged: markup may help grounding, with no promise of visibility or traffic.

JSON-LD sits in a script element with its type set to application/ld+json, which the specification calls a data block.⁵ The HTML standard says a script element does not represent content for the user, and that a data block is left to scripts or other tools to process.⁶ A fetcher that turns a page into readable text has every reason to discard script elements with the rest of the code. That is my inference from the formats, and the tests below suggest it is what happens.

The providers document very little about this step. Anthropic's web fetch tool retrieves the full text content of a URL and passes it to Claude as a plain-text document, does not support pages rendered with JavaScript, and does not say whether script blocks survive.⁷ Google's URL context tool for Gemini checks an internal index cache before falling back to a live fetch, which shows how easily the first two routes blur, and does not describe how it processes HTML.⁸ OpenAI's and Perplexity's crawler pages say nothing about how a page becomes text.⁹¹⁰

What the tests found

searchVIU, which makes SEO software, tested a fictional product page in October 2025 and published the results that December. Each of eight variants had its price in a different place, including visible text, JavaScript, JSON-LD, Microdata and RDFa. ChatGPT reported only prices in visible HTML, Gemini also found one added by JavaScript, and Claude reported none. Perplexity and Google AI Mode found nothing until the page was indexed, and one or two prices after. None reported a price that existed only in JSON-LD. The authors note that markup could still be used at indexing or in training, which a live test cannot measure.¹¹

Pilot Digital, a Chicago marketing agency, ran a more thorough version in August 2026 and published it in October. Ben Hicks hid 14 code words on a static page, each placed differently, and asked ChatGPT, Gemini and Claude to find them with their own fetch tools, alongside Firecrawl and Chrome's reader view. JSON-LD came through in none of 25 attempts, wherever it sat. The same JSON in a hidden div reached all three assistants, while plain JSON in a script element was dropped like the schema, so the tools appear to discard script elements whatever they hold. By its own account the test covered one page on one day, set to noindex, and could not see what GPTBot or ClaudeBot store.¹²

One test points the other way. Mark Williams-Cook, reported by Search Engine Roundtable in February 2026, put a made-up company's address only inside invalid JSON-LD, and both ChatGPT and Perplexity returned it. His reading was that the systems pick up whatever text is in the HTML, valid schema or not.¹³ The report does not say whether the page was fetched live or found in a search index. Either way it shows markup arriving as plain text, with no sign it was used as structured data.

So a fact that lives only in JSON-LD is at risk whenever an assistant reads your page live, and the fix costs nothing: put it in visible text as well. Markup injected by a tag manager has a second problem, because it only exists after JavaScript runs: Anthropic's fetch tool does not support JavaScript-rendered pages,⁷ and in searchVIU's test Gemini was the only assistant that picked up JavaScript content when fetching the page directly.¹¹ Checking what crawlers see on a JavaScript site covers how to test for that.

Common Crawl's raw WARC files hold the full HTML of each page, markup included. What reaches a model depends on how that HTML is turned into text. The authors of FineWeb, an open training dataset built from 96 Common Crawl snapshots, found Common Crawl's own text extracts kept too much boilerplate and menu text, so they extracted text from the raw HTML with the trafilatura library instead.¹⁴

Open datasets like FineWeb document their pipelines. The labs behind ChatGPT, Claude and Gemini do not publish how theirs treat a JSON-LD block, so nobody outside them knows whether your markup ever reached a model, or what it did if it did. Pilot Digital noted the same gap in its own test.¹² I would treat any confident claim about schema and training data as a guess.

The search engines that parse markup into an index say it helps them understand pages, and none promises citations for it. OpenAI, Anthropic and Perplexity, whose assistants fetch pages live, say nothing about on-page markup.

Most figures in circulation come from correlation studies, which tend to find more markup on cited pages. The obvious explanation is that well-run sites add markup and also do everything else well. Ahrefs found AI-cited pages almost three times as likely to carry JSON-LD across 6 million URLs, and raised that explanation itself.²⁰

Its follow-up is the closest thing to a controlled test at scale. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched pages that never did, comparing citations for 30 days either side. ChatGPT and AI Mode citations moved by amounts indistinguishable from zero, and AI Overviews citations fell slightly, which the authors could not attribute to the markup. Every page already had more than 100 AI Overview citations in February 2025, so the study cannot say whether markup helps a page nobody cites yet.²⁰

• SE Ranking's model of 216,524 pages cited by ChatGPT, published in November 2025, found that pages with FAQ schema averaged 3.6 citations against 4.2 for pages without, and concluded that schema alone does not significantly raise the chance of a citation.²¹
• Search Atlas, which sells SEO software, compared domain-level schema coverage with visibility in OpenAI, Gemini and Perplexity in a December 2025 study and found the distributions nearly identical at every level. It published no sample size or statistical tests.²²
• Kurt Fischman of Machine-like, a firm that sells AI visibility work, analysed 730 citations from ChatGPT (GPT-4o) and Gemini 1.5 Pro across 75 commercial queries in a preprint published in February 2026. Once errors were clustered by query, schema presence was not a significant predictor; Google organic rank was the strongest. Product and Review pages with populated prices, ratings or specifications were cited 61.7% of the time, against 41.6% for generic types such as Article or Organization.²³

In my view page type could explain that last gap as easily as markup, since a product page with a price answers a different kind of question. None of this proves markup is useless for AI answers. It does show that the uplift figures in sales pitches have no controlled evidence behind them, and the one large controlled test I could find measured no lift.

Product data

Shopping is where structured data does documented work. Google recommends both Product markup and a Merchant Center feed, and says markup helps it understand the page.²⁴ Bing says IndexNow tells it that something changed while structured data tells it what changed, and that merchants using both can skip a separate feed.²⁵ ChatGPT's shopping results draw on structured metadata from providers and on direct feeds.¹⁹ Keeping feed, markup and page in agreement is covered in the post on product data for AI shopping.

Business details that match your profiles

For a local business, the profile does more work than the markup. Google asks site owners to keep Business Profile and Merchant Center information up to date for its AI features,³ and Bing points local businesses to Bing Places so that address, hours and contact details stay current and eligible for AI answers.²⁶ LocalBusiness markup earns its place when it repeats those details exactly, and adds a contradiction when it does not. Your Google Business Profile is doing more work than your website explains why.

Organization markup and sameAs

Organization markup helps Google understand an organisation's administrative details and tell it apart from others, with sameAs pointing to pages elsewhere that carry more information about it.²⁷ Schema.org defines sameAs as a page that unambiguously indicates the item's identity, such as a Wikipedia page, a Wikidata entry or an official website.²⁸ Bing's guidelines say clear, consistent entity definition improves grounding visibility and citation accuracy.⁴ It is a small job that helps search engines work out which entity a page is about, and I would do it for any business.

• Implement one Organization node, or a LocalBusiness subtype for a single-location business, on the home or About page, with a stable @id, the legal name and sameAs links to profiles that are unmistakably yours.
• Implement Product and Offer markup on every product page, generated from the same record as your Merchant Center feed and any feed you send to ChatGPT.
• Implement markup that still earns rich results where the page really is that kind of page: breadcrumbs, articles, events, recipes, reviews of products.
• Keep accurate every value in the markup against the visible page, your Business Profile and your feeds. Google and Bing both say markup must match what visitors can see.³⁴
• Keep visible every fact you care about: prices, opening hours, addresses, phone numbers, credentials. A live fetcher may never see your JSON-LD.¹²
• Skip adding markup to already-cited pages in the hope of more citations. Ahrefs measured no lift.²⁰
• Skip FAQPage markup added for AI visibility. In SE Ranking's data, pages with FAQ schema averaged fewer ChatGPT citations.²¹
• Skip “AI schema” packages that promise citations. Google says there is no special schema.org markup to add for its AI features.³

<p>Example Dental Leeds is open Monday to Friday, 8:30am to 5:30pm.
Call us on 0113 496 0000.</p>

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Dentist",
  "@id": "https://www.example.com/locations/leeds/#location",
  "name": "Example Dental Leeds",
  "telephone": "+44 113 496 0000",
  "openingHours": "Mo-Fr 08:30-17:30",
  "parentOrganization": { "@id": "https://www.example.com/#organization" }
}
</script>

The business, number and URLs are placeholders. The fact sits in visible text, which a live fetcher keeps, and in markup, which an index can parse. Get the facts right and visible first, make your profiles and feeds agree with them, then add markup that describes them. Done in that order, markup does the jobs it is documented to do. Bought as a citation tactic on its own, it has no evidence behind it.