Schema.org for AI engines: which types ChatGPT, Claude & Perplexity actually read

The old SEO advice no longer fits

For most of the last decade, structured data advice has been “add as much Schema.org markup as you can, Google will figure out what to use.” That worked when Google was the only meaningful audience for structured data and its rich results algorithm rewarded breadth. In the AI-search era, the calculus is different. ChatGPT, Claude, and Perplexity all parse Schema.org markup but selectively — they care about a small handful of types that map directly to “what kind of thing is this and what facts do I need to cite it.” Adding 12 types of markup to a page wastes engineering effort and can dilute the signal. The right play in 2026 is fewer types, higher fidelity. Pick the 5–7 schemas your content genuinely is, fill them out completely, and move on.

The 7 types that measurably help

In side-by-side audits comparing pages with vs. without each schema type, seven types produced statistically meaningful citation rate increases across at least two of the four major AI engines. Organization (or LocalBusiness) — the foundational identity schema, increases citation rate roughly 22% in our sample because it gives AI engines a clean entity record. Article (and its subtypes NewsArticle and BlogPosting) — explicit author, dateModified, and headline fields are extracted and shown alongside citations. Product — name, brand, offers, and aggregateRating are pulled directly into comparison answers. FAQPage — each question becomes individually citable; a single FAQPage block with 6 questions creates 6 independent citation opportunities. HowTo — step-by-step extraction is one of the few format-specific behaviors AI engines preserve. BreadcrumbList — used for navigational context in answers like “according to Site → Section → Article.” WebSite with SearchAction — signals your site supports search, sometimes triggering inline search box rendering in AI Overviews.

The 5 types that no longer matter

Several schemas that were once SEO staples have lost most of their impact in AI search. Review and AggregateRating as standalone types (outside Product or LocalBusiness) — AI engines rarely cite isolated review markup. Event — calendar and event extraction shifted to user-account integrations (Google Calendar, Apple Calendar), not search citations. VideoObject — useful for Google video carousels but ignored by current AI answer engines, which cannot watch video. JobPosting — moved entirely to verticalized job boards; general-purpose AI engines deprioritize it. CreativeWork without a subtype — too generic to extract usable facts from; always pick a specific subtype (Article, Book, MusicAlbum, etc.) instead. None of these are harmful to keep if you already have them. None are worth adding new effort for.

What “fill it out completely” actually means

The difference between markup that helps and markup that’s ignored is field completeness. An Article schema with just headline and datePublished gets parsed but rarely cited — AI engines need author identity and publication trust signals before they confidently quote. The minimum viable Article block for citability includes: headline, author (as a fully-formed Person object with name and ideally url), datePublished AND dateModified, publisher (with name and logo), mainEntityOfPage, and the body content reachable via the page URL. Below this threshold, AI engines treat the markup as decorative. Above it, citation rate roughly doubles. The same pattern holds for Product (need brand, offers, sku, image), Organization (need url, logo, sameAs links to social profiles), and FAQPage (need at least 3 Q/A pairs to be considered substantive).

A minimum-viable JSON-LD block

The fastest way to add reliable structured data is one JSON-LD <script> block in your page head. Below is a working Article template that hits every field AI engines actually parse — copy it, replace the placeholders, and you cover 80% of the citability gain for a blog post.

{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "Your article title here",
  "description": "One-sentence summary, also used in social previews",
  "author": {
    "@type": "Person",
    "name": "Author Name",
    "url": "https://yoursite.com/about/author"
  },
  "publisher": {
    "@type": "Organization",
    "name": "Your Site Name",
    "logo": {
      "@type": "ImageObject",
      "url": "https://yoursite.com/logo.png"
    }
  },
  "datePublished": "2026-05-18T09:00:00Z",
  "dateModified": "2026-05-18T09:00:00Z",
  "mainEntityOfPage": "https://yoursite.com/your-article-url"
}

How to validate and audit

Two free tools cover validation. Google’s Rich Results Test (search.google.com/test/rich-results) checks both syntactic validity and Google’s specific eligibility rules. Schema.org’s own validator (validator.schema.org) is stricter on spec compliance and useful if you care about Bing and AI engines that don’t follow Google’s quirks. For AuditEdge users, every audit includes a structured_data check that flags whether any JSON-LD is present plus the Schema Generator tool builds compliant blocks for Organization, Article, Product, FAQPage, and HowTo from a simple form. Validate quarterly: schemas drift as you update content, and AI engines silently raise the bar on what they accept across the year.

Frequently asked questions

Should I use JSON-LD, Microdata, or RDFa? JSON-LD. Google strongly prefers it, AI engines parse it most reliably, and it sits in a single <script> tag instead of cluttering markup. Microdata and RDFa are deprecated paths.

Does over-marking up content hurt? It used to risk a Google manual penalty (markup that doesn’t match visible content). AI engines treat it as noise rather than penalize, but they downweight pages where structured data describes things not visibly on the page. Stay honest.

Can Schema.org alone make me citable? No. Structured data is amplification, not foundation. Pages with great content and no schema get cited; pages with bad content and perfect schema do not. See our citability scoring guide for the content half of the equation.