Structured Data Formats That Influence AI Answers
Labeled markup helps machines extract and trust content, though platform support remains uneven.

Structured data formats shape AI answers by making content easier to pull out, link to its source, and check. From JSON-LD markup to enterprise knowledge graphs, their value lies in replacing inference with labeled facts. That makes structured data a machine-readable trust signal. Its implementation should follow that purpose, while accounting for real uncertainty about how different AI platforms use it.
How AI systems turn page content into cited answers — and where structured data fits in
AI answer engines must extract meaning before they can synthesize a response. A page may contain useful information without making clear who supplied it, what kind of content it is, or which statements belong together. A string of numbers could be a price, a phone number, or a date. Context alone must then do the work. When an answer draws on many sources, those ambiguities complicate both extraction and attribution.
Structured data labels the elements that would otherwise require interpretation, including questions and answers, authorship, content type, and entity identity. Those labels give systems explicit claims to process rather than leaving every relationship implicit in prose. They make it easier to identify what the publisher claims and connect it to the source, which supports more confident extraction and citation.
What structured data actually is and how the Schema.org vocabulary became the shared standard
Structured data organizes information into fields defined by a fixed schema. Unstructured content, including emails, documents, and free-form text, leaves more of that organization for the reader or machine to infer. A recipe can explain its ingredients and cook time in prose while also labeling those elements for machines. A blog can similarly identify its author and publication date.
Schema.org supplies the shared vocabulary for those labels. Google, Microsoft, Yahoo, and Yandex launched the project together so systems could use the same semantic language for content types. The vocabulary names what the information represents. The delivery format determines how those labels reach machines, with JSON-LD, Microdata, and RDFa providing three ways to express schema markup.
Google prefers JSON-LD, which keeps structured data separate from the page’s visible HTML markup. That separation gives crawlers a distinct representation of the page’s content and makes implementation flexible. Microdata and RDFa remain valid alternatives for expressing schema. JSON-LD delivered in the document head remains the implementation standard after Google’s March 2026 core update, but choosing the format is only the first decision. The schema must also describe the right information.
The schema types that do the most work for AI answer generation
Schema.org includes more than 800 types, but AI answer generation gives particular importance to a much smaller group. Organization schema establishes the source’s identity through its name, URL, logo, and contact details. Identity comes before the decision to cite. Its sameAs properties can connect that identity to Wikidata, LinkedIn, and other authoritative directories, helping systems resolve references to the same organization across the web.
Article and BlogPosting schema identify editorial content, its author, and when it was last updated. Their dateModified property supplies a recency signal relevant to AI Overviews and AI Mode. Attribution adds another layer to that context. Pairing Article markup with Person schema tied to an author bio page gives systems more explicit information for assessing who stands behind the content. These types serve a different purpose from Organization markup, describing the work and its author rather than only the publisher.
FAQPage schema labels questions and their corresponding answers, making concise responses easier to isolate for summaries. Google took FAQ rich results out of Search on May 7, 2026. That was a change in display eligibility. It does not make the question-and-answer structure useless for machine extraction, although it removes the visible search feature that previously justified many implementations.
Review and AggregateRating schema supply quantified sentiment through scores and counts that can inform comparisons, particularly for local businesses and service providers. LocalBusiness schema serves a related need by showing which business a location-based answer refers to. Consistency is essential to that identification process. When LocalBusiness markup and Google Business Profile data disagree, those differences make the source harder to trust. These types therefore earn their place through the content and entities they describe.
Why the value of structured data shifted from visual search features to machine-level trust verification
For years, the most visible return on structured data was a richer search listing. Star ratings, FAQ dropdowns, and breadcrumb trails gave publishers an obvious display outcome to pursue. That rationale no longer covers the full opportunity. With FAQ and HowTo rich results retired, markup can retain machine-level usefulness even when it produces no special search appearance.
Microsoft has said that structured data helps its AI systems understand content. Google states that no special schema is required for AI Overviews or AI Mode, while allowing that structured data may benefit its systems generally. Markup can clarify content without being an admission requirement for AI answers, and clarification can support the knowledge-graph and context layers used in interpreting a source.
Reports of improved AI Mode citation rates for clean entity markup, and demotion where markup described supplementary rather than primary content, point toward alignment. Those observations reinforce the need to separate rich-result eligibility from entity trust, with different validation approaches and success measures for each. Schema can pass technical checks and still describe the wrong thing.
The unresolved question of whether AI systems parse JSON-LD semantically or read it as plain text
The strongest technical objection concerns what AI systems actually receive and process. Reading a JSON-LD block is not the same as interpreting it through Schema.org’s semantic rules. The distinction changes what a test can prove. Successful extraction may show that a model encountered the text, without showing that it validated or understood the markup as schema.
SEO researcher Mark Williams-Cook’s February 2026 test, with original credit to Richard Barrett, demonstrated that ChatGPT and Perplexity extracted an address placed exclusively inside invalid, made-up JSON-LD. The address did not appear in visible content. The result supports a narrow but useful conclusion. Those systems could read information from schema blocks as text, even when the markup could not supply a valid semantic structure.
Pilot Digital’s controlled test produced the opposite result at the retrieval stage. ChatGPT, Gemini, and Claude found none of the JSON-LD blocks in that test, with the explanation that the script tags containing them were discarded during fetching. The discrepancy cannot simply be dismissed as noise. It shows why claims about a platform “using schema” need to distinguish access to the block, extraction from its text, and semantic interpretation.
Platform and retrieval behavior are not uniform, and the evidence does not support a universal parsing model. Search-enabled systems may extract JSON-LD in circumstances where another fetch path omits it. That uncertainty limits the promise publishers can make. The case for implementation rests on the combined clarity markup provides to crawlers, knowledge graphs, and extraction systems.
How knowledge graphs extend structured data logic from web pages into enterprise AI systems
Knowledge graphs take labels beyond single pages and use them in networks of entities, relationships, and properties. Their nodes stand for things, and labeled edges show how those things relate. Relationships become part of the data itself. An AI system can therefore follow connections that raw document retrieval leaves implicit, rather than relying only on similarity between a question and a passage.
GraphRAG brings knowledge graphs and large language models together so models can reason across those connections. Microsoft GraphRAG is a named example of this approach. The link to page-level schema is conceptual. Both make meaning explicit, and an enterprise graph requires connected entities and relationships. That structure supports richer answers and traceable reasoning paths.
Data.world’s benchmark showed GPT-4 answering a substantially higher share of insurance questions correctly from a knowledge graph than from raw SQL of the same data. This comparison gives the enterprise argument concrete support. It shows that representation can materially change answer accuracy. The data alone is not the whole input.
The benefit also comes with an implementation trade-off. Conventional RAG deploys faster, while knowledge graphs require upfront entity extraction and relationship mapping. Where linked reasoning and accuracy are worth that work, a graph spells out relationships that similarity matching alone cannot supply. Where deployment speed dominates, the additional modeling work can constrain adoption.
llms.txt as a site-level complement to page-level schema
Page-level markup explains individual content items, but it does not curate a site’s preferred sources. Jeremy Howard of Answer.AI proposed llms.txt in 2024 to meet that need for the site as a whole. The convention resembles a hand-curated sitemap for AI assistants. Where sitemap.xml lists URLs, llms.txt identifies the pages a publisher considers authoritative and ranks them by importance, giving systems a more selective guide to the site.
As of May 2026, llms.txt remains a community convention outside the W3C and IETF standards. Thousands of sites serve /llms.txt, and adopters include Anthropic, Cloudflare, and Vercel. Adoption does not establish uniform behavior by AI platforms. The file expresses publisher guidance, but its presence alone does not demonstrate that an assistant follows the ranking or uses the nominated pages in answers.
Its role also differs from that of robots.txt. While robots.txt addresses crawler access, llms.txt addresses AI systems with curated information about available content and intended uses, including citation and training preferences. There is no guarantee that those preferences will be enforced. The relationship to schema is similarly complementary, with page markup describing what content means and llms.txt indicating which sources the publisher regards as authoritative. Neither layer substitutes for the other.
Making deliberate implementation decisions: what to prioritize and how to audit what you have
For practitioners, the implementation priority is entity clarity before markup breadth. Begin with Organization schema and sameAs identifiers that connect the organization to authoritative external sources. Then check those claims against visible content. Inconsistency between markup and the page undermines the trust signal the implementation is meant to provide, even when the syntax passes validation.
Schema selection should follow the page’s primary purpose. Apply FAQPage where FAQ content is primary, Article where editorial content is primary, and Product where product detail is primary. Google’s March 2026 update reinforced this alignment algorithmically, making opportunistic type selection a poor substitute for an accurate description of the main content.
Validation should keep going after markup is first added because changes to the content can make markup that was once correct become wrong. An ongoing audit needs to examine both detectability and correspondence with the current page. BrightEdge’s SearchIQ helps assess schema detectability and competitive implementation by identifying the schema types used on top-ranked pages for similar keywords.
For publishers adding llms.txt, the practical task is to identify and rank the most authoritative pages on the site’s core topics. Enterprise teams must weigh the accuracy and traceability they need against the work of preparing a knowledge graph and the faster deployment of RAG. Both decisions concern how much meaning to make explicit. Across page schema, entity markup, site-level guidance, and enterprise graphs, structure is useful because it replaces inference with labeled fact. The more a system must guess about meaning or identity, the weaker its basis for confident citation.


