Structured Data for LLMs: JSON-LD That Agents Actually Use

JSON-LD is the embedded format for Schema.org markup, a vocabulary launched in 2011 by Google, Bing, Yahoo and Yandex. Sources describe it as a machine-readable layer that ChatGPT, Claude, Perplexity and Gemini can consume to classify content, identify authors and extract citable claims. Claims about its influence vary. Microsoft Bing has confirmed it uses schema for its LLMs, one study finds schema on around 65% of pages cited by AI Mode, and Google says structured data does not directly influence rankings. This brief covers which types to deploy, how to map them to templates, how to validate them, and where the evidence is weak, so developers and agents can implement it with realistic expectations.

Key concepts

Schema.org

An open vocabulary of machine-readable types such as Article, FAQPage, HowTo, Organization, Person and DefinedTerm, launched in 2011 by Google, Microsoft Bing, Yahoo and Yandex. S1 · alicelabs.aiS4 · walkersands.com

JSON-LD

JavaScript Object Notation for Linked Data: a lightweight text format that conveys entities and their relationships and can be added to a page without altering its HTML. S2 · szymonslowik.com

FAQPage markup

Question entities with a name and an acceptedAnswer of type Answer with text. Self-contained answers give the highest extraction rate for AI citation, according to S1. S1 · alicelabs.ai

Shared Organization @id

Referencing the same Organization @id across all pages lets LLMs disambiguate the publisher entity, which is defined with name, url, logo and sameAs. S1 · alicelabs.ai

Generative Engine Optimization

GEO is the practice of influencing how AI systems interpret and reference a brand, with schema markup as one supporting method. S4 · walkersands.comS8 · digidop.com

Deploying JSON-LD across a site

1. Audit existing markupRun Google's Rich Results Test on 10-20 representative URLs, including homepage, pillar pages, articles, product pages and the About page. Note which types are present, missing or malformed, and build on CMS-generated markup.
2. Map types to templatesAssign schema types per page template, not per URL: Article for posts, DefinedTerm for glossary entries, HowTo for guides, Organization plus Person for About pages. Choose the smallest set covering 95% of content.
3. Write the JSON-LD fieldsInclude required and recommended properties. For Article: headline, author, publisher, datePublished, dateModified, image, mainEntityOfPage. Reuse one Organization @id site-wide.
4. Validate the markupCheck with Google Rich Results Test, validator.schema.org, Bing Markup Validator and Search Console coverage reports. Confirm schema content, such as HowTo step counts, matches what appears on the page.
5. Monitor and maintainReview Search Console structured data reports after deployment and keep dateModified current, since freshness signals are weighted heavily by LLM retrieval systems according to S1.

Schema types worth deploying

Map schema types to page templates rather than individual URLs. S1 recommends the smallest set that covers 95% of content, typically Article, FAQPage, HowTo, Organization, Person, DefinedTerm and BreadcrumbList. Article needs headline, author as Person, publisher as Organization, datePublished, dateModified, image and mainEntityOfPage. Both dates should always be present because freshness signals are weighted heavily by LLM retrieval systems. FAQPage entries use a Question with a name and an acceptedAnswer of type Answer with text. Answers of 2-4 sentences stay quotable in isolation, and five to eight per pillar page is described as the practical sweet spot. HowTo needs HowToStep entities with name and text, and step counts must match the page. Organization with name, url, logo and sameAs builds the publisher entity, and reusing the same Organization @id across pages helps LLMs disambiguate it. Person markup on author bios ties each article to a real human with credentials.

What the evidence says about LLM use

S7 reports that in March 2025 Fabrice Canel of Microsoft Bing confirmed at SMX Munich that schema markup helps Microsoft's LLMs understand content, specifically for Copilot. S6 found that around 65% of pages cited by AI Mode and 71% cited by ChatGPT use schema markup. The most common types on AI-cited pages were BreadcrumbList (38–42%), WebSite (34–36%), Organization (31–32%) and WebPage (31–36%). Most AI-cited pages use just 1–2 types (43% for AI Mode, 42% for ChatGPT). S6 also reports that structural schemas such as BreadcrumbList or WebPage show weak negative correlations with AI citations, while Article shows a weak positive one. It calls the effects tiny and says clear, well-structured content matters most. S8 notes John Mueller of Google confirmed in 2025 that structured data does not directly influence rankings. S2 and S3 argue more strongly that LLMs absorb JSON-LD as a data layer, a view S3 ties to reasoning models, retrieval architectures and knowledge graphs.

Validation and LLM-generated markup

Validate before and after deployment. S1 lists Google Rich Results Test, validator.schema.org and Search Console structured data coverage reports, and S6 adds Bing Markup Validator. Keep schema consistent with visible page content, and avoid markup such as third-party reviews or inactive job postings, which S6 says can put SEO and AI visibility at risk. Generating markup with LLMs needs the same care. S5 reports that around 75% of web pages still lack Schema.org markup, and that 40%–50% of markup produced by GPT-3.5 and GPT-4 is invalid, non-factual or non-compliant with the Schema.org ontology, which has 806 types and over 1,400 properties. Specialized LLM-powered agents that filter these errors improved quality, and after filtering GPT-4 outperformed humans. For agents that write JSON-LD, a validation and filtering step is therefore part of the workflow, not an optional extra.

Classical SEO schema vs LLM-oriented JSON-LD

Timeline

1. 2011
   Schema.org launched by Google, Bing, Yahoo and Yandexwalkersands.com
2. 2025-03
   Bing's Fabrice Canel confirms schema helps Microsoft LLMs at SMX Munichquoleady.com
3. 2025
   Google's John Mueller says structured data does not directly influence rankingsdigidop.com
4. 2025
   LLM4Schema.org benchmark published, assessing LLM-generated Schema.org markupjournals.sagepub.com

Common pitfalls

• Schema that does not match page content, such as HowTo step counts differing from the visible steps, is a common Rich Results Test failure.
• LLM-written markup is unreliable unfiltered: 40%–50% of GPT-3.5 and GPT-4 output was invalid, non-factual or non-compliant with the ontology.
• Do not over-credit schema: structural types showed weak negative correlations with AI citations, and S6 calls the effects tiny.
• Markup for third-party reviews or inactive job postings can put SEO and AI visibility at risk.

Key takeaways

• Map a small set of validated schema types to page templates, and keep datePublished, dateModified and a shared Organization @id consistent.
• Evidence of LLM use is confirmed for Bing and correlational elsewhere; Google says rankings are not directly affected.
• Validate all markup, especially anything generated by an LLM, and keep it consistent with visible page content.

FAQ

How this page is made. Built from the full text of 8 sources found with the AISA Tavily API (advanced search), refreshed weekly. A language model organizes what those sources state into the blocks above; each block links the sources it draws on, and figures that do not appear in the sources are removed.