Schema for AI search: what actually helps

Adding schema markup is unlikely to get a page cited by AI assistants, on the evidence available. The closest thing to a causal test found no lift in ChatGPT or Google’s AI Mode, and Google says its AI features need no special markup. Accurate schema still helps with Google’s rich results and with stating who you are, so keep it for those reasons.

Quotable offers a free schema markup generator and schema checker, even though the evidence that schema helps with AI search is weak. They help you write valid markup and find broken blocks. Neither tool will make an assistant cite or name you.

Does adding schema make AI assistants cite you?

No published test shows that adding schema causes AI citations, and the best one found no lift. Larger correlational studies disagree with each other, depending on how they compare pages. The kind of evidence matters more here than the headline number.

The Toronto result shows why simple comparisons mislead. Schema looks helpful between websites because that comparison credits the markup with the strengths of the whole site. Compare pages on the same site and the effect disappears. The authors call this an instance of Simpson’s paradox. In their niche B2B market the within-site figure was +16.2%, but it passed only two of five robustness tests, so we wouldn’t build on it.

SISTRIX and Toronto both compare within a site and still disagree. We don’t know why. The two differ in region, market and engines, and a page with Article markup may simply be a different kind of page. SISTRIX’s own warning is worth keeping in mind: its FAQ schema result moved from 23% less common at 833 URLs to 31% more common at 4,600, and 19% more common at 96,418.

Can AI assistants read JSON-LD at all?

Live fetches may not treat JSON-LD as structured data at all, though the tests are small. Two small practitioner tests hid a fact in markup and asked assistants for it, and they appear to disagree. Neither says anything about what happens to markup inside a search index.

• searchVIU, an SEO software company, tested ChatGPT, Claude, Gemini, Perplexity and AI Mode on 30 October 2025. It used one page with eight price placements. No system gave a price that existed only in JSON-LD. Claude gave no prices at all, even visible ones.
• Mark Williams-Cook, an SEO practitioner, reported in February 2026 that ChatGPT and Perplexity returned an address found only in deliberately invalid JSON-LD. He concluded they were reading it as ordinary page text.

Together these suggest a practical rule, which is our inference and not a finding: put every fact you care about in visible text. If a system reads your HTML as text, visible text reaches it. If it strips the markup, visible text still reaches it. Don’t copy these tests on a real site, because Google’s policies forbid marking up content readers can’t see.

Lab work adds little. A March 2026 preprint from authors at WordLift, which sells AI search software, found a small accuracy gain from adding JSON-LD to plain HTML in a retrieval pipeline built on Vertex AI. It didn’t test live pages or live assistants.

What do Google, Microsoft and OpenAI say about schema?

The operators’ statements differ, and none of them promises citations. Each one describes how that company’s own system is meant to work, not what any page will get.

• Google says structured data isn’t required for generative AI search and no special schema exists for it. It still recommends structured data for rich results. Its AI features page asks that markup match the visible text.
• Microsoft is warmer. A Bing product manager wrote in October 2025 that schema helps search engines and AI systems understand content. That post gives no measurement, and it appeared on Microsoft’s advertising blog.
• OpenAI’s crawler page doesn’t mention structured data. For products, it documents a separate product feed in its own format, open to approved partners. So page markup and being listed in its shopping results are different jobs.

Which schema types are worth having?

Most tech companies and brands need only a few schema types, chosen for what they do in Google rather than for AI. We checked these against Google’s search gallery and change log on 6 November 2026.

FAQPage shows how little the correlations tell you. SISTRIX found it 1.5 times more common on cited pages and stresses that the questions must be real. In the Toronto study, FAQ blocks came out at −3.4% within websites, not significant. That was the visible block, not the markup, so the two can’t be set against each other directly. Either way, nobody has shown that adding it causes citations.

What should you do about schema for AI search?

Treat schema as a small piece of housekeeping and do it in this order:

A minimal example for a software company, with both blocks on the relevant pages:

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

<script type="application/ld+json">
  {
    "@context": "https://schema.org",
    "@type": "BlogPosting",
    "headline": "How log retention affects your audit",
    "image": ["https://example.com/images/retention.png"],
    "datePublished": "2026-11-06T09:00:00+00:00",
    "dateModified": "2026-11-06T09:00:00+00:00",
    "author": {
      "@type": "Person",
      "name": "Sam Example",
      "url": "https://example.com/about/sam"
    }
  }
</script>

Google recommends putting only the person’s name in author.name, with no job title or publisher name.

What can go wrong with schema?

The main risk is markup that says something the page doesn’t. Google’s policies say markup must describe visible content and truly represent the page. A breach can bring a manual action, which removes eligibility for rich results but doesn’t affect ranking in web search. Google also added a guideline on fake and undisclosed incentivised reviews in July 2026. Valid markup isn’t a promise either, since Google doesn’t guarantee a feature will appear even after the page passes its test.

What our tools and data can’t show

Our generator and checker tell you whether markup is valid and which types can earn a Google rich result. They can’t tell you whether markup changes what an assistant says about you. Nobody can promise that, including us.

You could try your own test, because schema is set page by page. Add it to a random half of similar pages, keep the other half as controls, and track citations on both over several weeks. Be realistic about the result. Ahrefs used 1,885 pages to look for small effects, so a small site’s test will probably be inconclusive. Quotable sells a tool that tracks AI answers, and it has that same limit.

For most teams, retrieval is the bigger question. In the Toronto preprint, whether a page came back from ChatGPT’s own searches explained 32% of the within-website variation in citations, against 2.8% for all page features. It’s one correlational study of API models in narrow markets, not yet peer reviewed. Our guide to AEO covers that part.