Structured Data for AI Search: The 2026 Schema Playbook for LLM Citation

Classic SEO treated schema as a rich-result enhancement: ship FAQPage, get the accordion-style SERP feature, win the click. The AI search model uses schema for a different job. The LLM reads structured data in three layers:

• Knowledge graph ingestion at training time. Major LLMs (Claude, GPT, Gemini, Perplexity Sonar) crawl the public web and ingest JSON-LD into an internal entity graph. Pages with schema get a richer node: typed entity, known relationships (author, publisher, mentions), and verifiable claims. Pages without schema get a flat text node with weaker citation lineage.
• Answer-shape parsing at runtime. When a buyer asks a question, the LLM does not just retrieve documents. It assembles a typed answer. FAQPage and HowTo schemas tell the LLM the answer is already pre-shaped for the question, so the LLM lifts the schema text almost verbatim.
• Entity disambiguation.sameAs links on Organization and Person schemas tell the LLM "this brand on forkoff.xyz is the same brand on LinkedIn, Crunchbase, GitHub." Without that link, the LLM may merge or split the entity wrong and cite the wrong brand entirely.

A clean SERP rich-result strategy from 2022 still helps in 2026, but it is the floor, not the ceiling. The AEO floor adds Dataset, ItemList, dateModified discipline, and sameAs entity wiring. See the AEO strategy guide for the broader measurement layer that sits on top.

The schema.org vocabulary lists hundreds of types. Five carry the bulk of LLM citation weight in 2026. Each section below ships a copy-pasteable example. Validate every graph in the Schema Markup Validator before shipping.

• Article (or BlogPosting). Every long-form page on the site. Required for the LLM to treat the page as a citeable source.
• FAQPage. The highest-yield single add. Q + A blocks get lifted into answers verbatim.
• HowTo.Procedural content. ChatGPT cites HowTo steps almost verbatim on "how do I X" queries.
• Dataset. First-party data, research reports, benchmarks. The schema that drives /research and /stats citation.
• ItemList. Ranked or comparative content. Vendor comparisons, top-N lists, ranked playbooks. Tells the LLM the order is editorial, not arbitrary.

The cross-schema baseline that ships on every page regardless: Person + Organization + WebSite + BreadcrumbList. Covered in the cross-schema patterns section below.

Article (and its subtype BlogPosting) is the baseline schema for every long-form page on the site. Without it, the LLM treats the page as anonymous text. With it, the LLM resolves the page to a known author, publisher, and timestamp.

Required fields: headline, author (Person or Organization with @id), datePublished, image. Without all four, the page gets dropped from AI Overview eligibility and demoted in Perplexity citations.

Recommended fields: dateModified (Claude weighs this heavily), mainEntityOfPage, publisher (Organization with logo), description, articleSection. The reviewedBy field with a Person @id is the single strongest E-E-A-T add for LLM citation.

{
  "@context": "https://schema.org",
  "@type": "Article",
  "@id": "https://forkoff.xyz/guides/structured-data-for-ai-search#article",
  "headline": "Structured Data for AI Search: The 2026 Schema Playbook for LLM Citation",
  "description": "The schema playbook for LLM citation in 2026. Article, FAQPage, HowTo, Dataset, ItemList. Copy-pasteable JSON-LD for ChatGPT, Claude, Perplexity, AI Overviews.",
  "author": {
    "@type": "Person",
    "@id": "https://forkoff.xyz/#cofounder-simba",
    "name": "Kartik Chugh",
    "jobTitle": "Cofounder, FORKOFF",
    "sameAs": [
      "https://www.linkedin.com/in/kartikchugh123",
      "https://x.com/0x0simba_eth"
    ]
  },
  "reviewedBy": { "@id": "https://forkoff.xyz/#founder-kshitij" },
  "publisher": {
    "@type": "Organization",
    "@id": "https://forkoff.xyz/#organization",
    "name": "FORKOFF",
    "logo": {
      "@type": "ImageObject",
      "url": "https://forkoff.xyz/assets/landing/logo.png"
    }
  },
  "datePublished": "2026-06-09",
  "dateModified": "2026-06-09",
  "image": "https://forkoff.xyz/og/guides.webp",
  "mainEntityOfPage": {
    "@type": "WebPage",
    "@id": "https://forkoff.xyz/guides/structured-data-for-ai-search#webpage"
  }
}

Common shape errors that get pages dropped:

• String author instead of Person object. "author": "Kartik Chugh" validates as a string, not a typed entity. The LLM has nothing to link to. Always render author as Person with @id and sameAs. The sameAs array is the entity-disambiguation signal that gets the author resolved to a real human across the open web.
• Missing dateModified. Without it, Claude treats the page as ageless and demotes it on time-sensitive queries. When the content meaningfully changes, bump dateModified; do not bump on a typo fix or a cosmetic tweak, because LLMs cross-check modification frequency against content delta and de-rank pages that game the signal.
• Image as relative URL. "image": "article-slug.webp" (relative path) fails validation. Use absolute URLs in JSON-LD always. The same rule applies to publisher.logo, author.image, and every other image field across the graph.
• publisher.logo without dimensions. Google Rich Results Test requires publisher.logo to be an ImageObject with an explicit url. Render the logo as an ImageObject node, not a bare URL string, to keep the page eligible for the publisher- branded Top Stories carousel and for Bing News indexing.
• articleSection mismatched with the URL path. When articleSection says "Guides" and the URL lives at /blog/, the LLM treats the page as miscategorized and demotes it in cluster citations. Keep articleSection aligned with the breadcrumb leaf one level up.

FAQPage is the single highest AEO-yield schema type in 2026. Every Question + acceptedAnswer pair becomes a pre-shaped, citation-ready snippet that LLMs lift directly into answers. Pages with 5 plus FAQ entries cite at roughly 2 to 3 times the rate of pages without (per the FORKOFF audit ledger across 28 client engagements).

The shape that works:

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "What is structured data and why does it matter for AI search?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Structured data is JSON-LD markup that describes a page's content in machine-readable schema.org vocabulary. For AI search, it feeds the knowledge graph an LLM consults at training time and the answer-shape parser at runtime. Pages with valid schema get pulled into AI Overviews, Perplexity citations, and ChatGPT answers disproportionately."
      }
    },
    {
      "@type": "Question",
      "name": "Which schema type drives the most LLM citations?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "FAQPage. Every Question + acceptedAnswer pair is a citation-ready snippet that LLMs lift almost verbatim. Pages with 5+ FAQPage entries cite at 2 to 3 times the rate of pages without."
      }
    }
  ]
}

Common shape errors:

• Boilerplate questions."What is X?" with a 1-line answer scraped from the H1 cites poorly. Write substantive 40-to-120-word answers that actually answer the question.
• Missing acceptedAnswer.text. The Answer node must have a text property. Some CMS templates ship Answer with just an @type and no text. Validates as a structure violation; LLM drops the entry.
• Duplicate FAQPage on the same canonical URL. Rendering two FAQPage blocks on one page (e.g. a global FAQ component and a page-specific one) causes Google Search Console to flag "Duplicate field FAQPage" and demotes the page. Merge into one FAQPage with all questions.
• 5-question minimum (FORKOFF rule). FAQPage with fewer than 5 entries gets dropped from the V29 preflight gate. Either add real questions or remove the schema.

HowTo is the right schema when the user's query is procedural: "how do I add JSON-LD to a Next.js page", "how do I validate schema before shipping", "how do I structure a /vs page". ChatGPT cites HowTo steps almost verbatim on these queries, often beating a longer Article on the same topic because the answer is already pre-shaped.

When HowTo wins over Article for AI citation:

• The page describes a step-by-step process with a clear outcome.
• The buyer's query starts with "how do I" or "how to".
• Each step has a discrete action verb and a verifiable outcome.

The shape that works:

{
  "@context": "https://schema.org",
  "@type": "HowTo",
  "name": "How to add structured data to a Next.js page for AI search",
  "description": "Ship JSON-LD on a Next.js App Router page that passes Schema Validator and earns LLM citations.",
  "totalTime": "PT15M",
  "step": [
    {
      "@type": "HowToStep",
      "position": 1,
      "name": "Draft the Article JSON-LD object",
      "text": "Define the Article node with headline, author (Person with @id and sameAs), datePublished, dateModified, image (absolute URL), and mainEntityOfPage. Render reviewedBy with a second Person @id for E-E-A-T."
    },
    {
      "@type": "HowToStep",
      "position": 2,
      "name": "Inject the script tag in the Page component",
      "text": "Render <script type='application/ld+json' dangerouslySetInnerHTML={{ __html: safeJsonLd(jsonLd) }} /> at the top of the page return. Server-rendered so the LLM crawler sees it on first fetch."
    },
    {
      "@type": "HowToStep",
      "position": 3,
      "name": "Validate in Schema Markup Validator",
      "text": "Paste the rendered page URL into validator.schema.org. Fix every error (missing required field, wrong type, malformed URL). Warnings are advisory; errors block AI citation."
    },
    {
      "@type": "HowToStep",
      "position": 4,
      "name": "Re-validate in Google Rich Results Test",
      "text": "Confirm Google parses the same graph. Rich Results Test is stricter on Article requirements (image dimensions, publisher.logo) and catches issues the Schema Validator misses."
    }
  ]
}

Common shape errors: missing position numbers, missing step.text (just an image), and using HowTo for non-procedural content (a list of tips, an opinion piece). Wrong schema is worse than no schema; the LLM drops the page from procedural-intent answers entirely. The step.image field is optional but adds materially to rich result eligibility on Google; render it as an absolute URL on steps where a screenshot or diagram clarifies the action.

totalTime is a citation lever.Buyers asking procedural questions often filter by time-to-complete. Rendering totalTime as an ISO 8601 duration (PT15M for 15 minutes, PT2H for 2 hours) lets the LLM surface the page on "quick" or "under 30 minutes" modifier queries. Missing totalTime leaves the page invisible to those modifier filters.

Dataset is the schema that drives citation for research, benchmarks, and first-party data pages. Underused by 90 plus percent of marketing sites, which is why a single well-marked Dataset page can out-cite a competitor's entire blog on the same topic.

The use case: any page that publishes a number, a benchmark, a ranking, a study, or a dataset. ICP fit research pages, benchmark reports, survey results, audit ledgers, and the /stats hub all qualify.

The shape that works:

{
  "@context": "https://schema.org",
  "@type": "Dataset",
  "name": "FORKOFF AEO Citation Lift Benchmark 2026",
  "description": "Citation share lift across ChatGPT, Claude, Perplexity, and Google AI Overviews on a 150-query bank, measured weekly across 28 client engagements over a 12-month window.",
  "url": "https://forkoff.xyz/research/aeo-citation-lift-benchmark-2026",
  "creator": {
    "@type": "Organization",
    "@id": "https://forkoff.xyz/#organization",
    "name": "FORKOFF"
  },
  "datePublished": "2026-06-09",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "keywords": ["aeo", "citation rate", "llm citation", "ai search"],
  "distribution": [
    {
      "@type": "DataDownload",
      "encodingFormat": "text/csv",
      "contentUrl": "https://forkoff.xyz/research/aeo-citation-lift-benchmark-2026.csv"
    }
  ]
}

Required fields: name, description, creator, datePublished. Without all four, Google flags Dataset as ineligible for Dataset Search and the LLM treats the page as an unverified claim.

The distribution field is the moat. Linking a downloadable CSV or JSON of the raw data via DataDownload signals that the claim is verifiable. LLMs weight verifiable claims higher than unsourced ones. The CSV does not have to be the full dataset; a representative sample is enough.

The license field matters for AI Overview surfaces. Google AI Overviews preferentially cite open-licensed datasets because the citation policy is downstream of the data license. A Dataset with a Creative Commons license cites at materially higher rates than one with no license declared. The same logic applies on Perplexity, where the citation transparency layer reads the license field. Use CC-BY 4.0 unless the data is genuinely proprietary.

variableMeasured is the entity-graph add.For benchmark and survey datasets, the variableMeasured field enumerates the columns or metrics. Tells the LLM what the dataset can answer questions about. A benchmark on AEO citation rate with variableMeasured set to "citation share, source mention rate, answer position" gets cited on each of those three sub-queries independently.

ItemList is the right schema for any page that ranks or compares items: top-N lists, vendor comparisons, ranked playbooks, /vs and /best pages. Tells the LLM the order is editorial, not arbitrary, which is critical when the LLM is composing a vendor-list answer from your page.

The shape that works:

{
  "@context": "https://schema.org",
  "@type": "ItemList",
  "name": "Top 5 schema types for AI search citation in 2026",
  "itemListOrder": "https://schema.org/ItemListOrderAscending",
  "numberOfItems": 5,
  "itemListElement": [
    {
      "@type": "ListItem",
      "position": 1,
      "name": "Article",
      "url": "https://forkoff.xyz/guides/structured-data-for-ai-search#article-schema"
    },
    {
      "@type": "ListItem",
      "position": 2,
      "name": "FAQPage",
      "url": "https://forkoff.xyz/guides/structured-data-for-ai-search#faqpage-schema"
    },
    {
      "@type": "ListItem",
      "position": 3,
      "name": "HowTo",
      "url": "https://forkoff.xyz/guides/structured-data-for-ai-search#howto-schema"
    },
    {
      "@type": "ListItem",
      "position": 4,
      "name": "Dataset",
      "url": "https://forkoff.xyz/guides/structured-data-for-ai-search#dataset-schema"
    },
    {
      "@type": "ListItem",
      "position": 5,
      "name": "ItemList",
      "url": "https://forkoff.xyz/guides/structured-data-for-ai-search#itemlist-schema"
    }
  ]
}

Position discipline is the gate. The most common ItemList error is duplicate or non-sequential position values: two entries with position 1, or skipping from position 2 to position 4. Both cause Google Rich Results Test to fail and the LLM to drop the ranking signal. Positions must be 1..N, unique, sequential.

Cluster-relative positions break. When generating ItemList across paginated sets (page 2 of a ranking), some teams render positions 1..10 again instead of 11..20. The schema becomes ambiguous and the LLM treats the page as a fresh top-10, not the continuation. Always render absolute positions across the full ranking.

Three schema nodes ship on every page on the site regardless of content type. They are the entity backbone the LLM uses to disambiguate the brand:

• Organization (with sameAs).The brand entity. sameAs links to LinkedIn, Crunchbase, GitHub, X, Wellfound, Featured.com. Tells the LLM "these profiles are the same entity." Without sameAs, the LLM may merge the brand with a same-named entity or split it across profiles.
• WebSite (with publisher reference). The site entity. Links to Organization via publisher.@id. Tells the LLM the site belongs to the brand. Required for Sitelinks Search Box eligibility.
• Person (for author and reviewer). The human entity behind the content. Each Person node has @id, name, jobTitle, and sameAs. Article.author and Article.reviewedBy both point at Person @ids. E-E-A-T floor.

The fourth always-present node is BreadcrumbList. Renders the page's position in the site hierarchy. LLMs use it to understand which pages cluster together, which is critical for vendor-list answers where the LLM may cite the hub and the spoke as evidence of topical depth.

The mainEntityOfPage pattern.Every typed entity (Article, HowTo, Dataset) should set mainEntityOfPage to the WebPage @id. Tells the LLM "this is the canonical entity for this URL." Without it, the LLM may pick the wrong node when the page renders multiple typed entities.

For the broader entity-disambiguation play see the agent-ready site audit, which covers the sameAs ledger and entity-graph audit FORKOFF runs on every engagement.

Schema that validates locally still fails in production if the deployment introduces escaping bugs, missing fields, or stale dateModified. The FORKOFF ship pipeline runs four validators sequentially. Skipping any of them ships broken schema.

1. JSON-LD Playground (author-time). json-ld.org/playground catches malformed JSON, missing @context, and circular @id references. Run this on the raw JSON before pasting into source.
2. Schema Markup Validator (pre-deploy). validator.schema.org catches schema.org vocabulary errors: wrong @type, missing required field, invalid property on a type. Errors block AI citation; warnings are advisory.
3. Google Rich Results Test (pre-deploy). search.google.com/test/rich-results is stricter than the Schema Validator on Article (image dimensions, publisher.logo) and FAQPage (Q + A count). Catches issues the Schema Validator misses.
4. Post-deploy GSC URL Inspection.After deploy, run the URL through Google Search Console's URL Inspection tool. Confirms Google parses the same schema graph it would for indexing. Catches Vercel-side caching or CDN-transform bugs that change the rendered HTML.

On the FORKOFF Website repo this pipeline is automated. The V29 preflight gate (audit-schema-completeness) runs on every staged commit, enforcing FAQPage 5-plus minimum, BreadcrumbList presence, and Article author @id wiring. The V36 gate audits tool embed authenticity. Failing V29 blocks the commit.

To audit a live site without the FORKOFF preflight, the FORKOFF AEO Checker runs the equivalent gate set on any URL and ships a schema-coverage scorecard with the specific failures and the patch.

This guide is the technical reference. The strategic and operational context lives on these adjacent pages:

• The AEO strategy guide covers measurement, the 30-day audit sprint, and per-LLM calibration. Pair with this guide for the full stack.
• The AI search ranking factors guide covers the 4 signal families and per-engine pipelines that this schema playbook plugs into.
• /services/answer-engine-optimization is the dedicated AEO engagement when implementation needs an operator.
• /services/answer-engine-optimization is the broader AI search operating model that wraps AEO + schema + measurement.
• /services/perplexity-seo is the Perplexity-first lane where Dataset and ItemList schema drive disproportionate citation lift.
• The agent-ready site audit covers schema as part of the broader site-readiness checklist (llms.txt, sameAs ledger, entity graph).
• The 2026 AEO playbook covers the operational sprint shape and the founder-facing decisions.
• How to get cited by ChatGPT in 2026 covers the ChatGPT-specific browsing-tool and recency signals.
• Best AI visibility tools vs FORKOFF methodology covers the tool-stack comparison for buyers evaluating vendors.
• How AI Overviews rank brands covers the Overviews-specific selection signals and the schema role inside them.

To audit an existing site against this playbook, the AEO Checker ships a schema-coverage scorecard with per-page failures and the patch.

The citation-lift figures (2 to 3 times for FAQPage-marked pages, 5-plus internal links per page as the source-authority floor, 90 plus percent of marketing sites underusing Dataset schema) come from the FORKOFF audit ledger, a private dataset built across 28 client engagements over the 12 months ending May 2026. Each engagement tracks per-LLM citation rate on a fixed query bank weekly. The aggregate numbers are operator-observation across that proof, not a peer-reviewed study.

Reproducibility notes: the lift figures are sensitive to baseline schema completeness on the site (a site with zero schema lifts more per FAQPage add than a site with partial schema), domain authority (cold-start sites lift slower), and query-bank composition (vendor- list queries lift faster than category-defining ones). Treat the figures as ranges, not point estimates. The methodology lives in the AEO service page.

The Schema Markup Validator and Google Rich Results Test references are vendor-neutral public tools at validator.schema.org and search.google.com/test/rich-results respectively. The validator stability has been stable across the 12-month observation window; schema.org vocabulary version is 27.0 as of June 2026.