Schema Markup Engineered for LLM Citation

Traditional Google ranking infers meaning from content; schema adds confirmation. LLMs in retrieval mode have a harder problem: from a corpus of millions of candidate passages, they need to select the few that best answer a specific prompt and then attribute them correctly. Schema markup provides explicit, structured hints that materially improve both the retrieval (this passage is about X) and the attribution (the author is Y, the publisher is Z, this is a definition versus an opinion).

The empirical pattern across our engagements: pages with full Article + Person + Organization schema get cited in AI assistants at roughly 1.5–2.5x the rate of equivalent-content pages with only basic Article markup. The marginal cost of deploying full schema is small; the citation-rate lift is consistent enough to justify it across content libraries.

The most consistently effective baseline is full Article schema with embedded Person (author) and Organization (publisher) — and crucially, sameAs properties on both that link to the entity graph. Article alone is table stakes. The full pattern signals 'this is content authored by a recognised person on behalf of a recognised organisation' — exactly the kind of source LLMs preferentially cite.

• Article: headline, datePublished, dateModified, articleSection, wordCount, mainEntityOfPage, image.
• author (nested Person): name, jobTitle, url (to author bio page), sameAs (LinkedIn, Wikipedia if applicable, ORCID for academics, X/Twitter, professional association profiles).
• publisher (nested Organization): name, logo (ImageObject), url, sameAs (Wikidata, Wikipedia, LinkedIn, Crunchbase).
• @id on the Article: a stable URI for the article entity, enabling cross-reference from other schema on your site.

FAQPage and HowTo schema map directly onto the answer shapes LLMs produce — question/answer pairs and step-by-step instructions. When an LLM is generating a response that calls for a list of FAQs or a how-to sequence, pages that have already structured their content in those exact shapes are disproportionately likely to be retrieved as the source.

The deployment pattern: every FAQ section on every page should be wrapped in FAQPage schema, even when there are only 2–3 questions. Every step-by-step process should be wrapped in HowTo schema with proper HowToStep nesting. The marginal authoring cost is near zero (most CMSes can template this); the citation-rate lift is reliably observable in LLM citation logs after 60–90 days.

Speakable schema (SpeakableSpecification) marks specific passages as 'safe to read aloud' — originally designed for Google Assistant voice responses. In 2025–2026 it has acquired a broader role as an explicit signal that the marked passage is a clean, self-contained, citation-ready answer. LLMs increasingly preferentially extract speakable-marked passages over surrounding context when generating cited responses.

The deployment pattern: identify the 1–3 sentences on each page that most cleanly answer the page's primary question, wrap them in Speakable cssSelector references in the page schema. The selected passages should be the same passages you'd put in a TLDR or summary box — definitional, complete, and quotable without surrounding context.

DefinedTerm schema (with optional DefinedTermSet for grouped glossaries) is heavily underutilised. It marks a specific term and its definition in a way that LLMs can directly retrieve when generating definitional answers. The pattern works best for glossary pages, in-content definitional asides, and any content that includes 'X is defined as Y' constructions.

Across our engagements with sites that have built out structured glossaries with DefinedTerm markup, those glossary pages become disproportionately frequent citation sources in AI assistants when users ask 'what is [term]' questions — often outpacing the site's deeper conceptual content on the same topic.

Dataset schema is the right markup for any original research, survey results, benchmark data, or proprietary statistics published on your site. LLMs are trained to identify and cite primary research, and Dataset markup explicitly signals 'this is primary data, not commentary on someone else's data.'

For sites that produce original research even occasionally, the citation lift from Dataset markup on those pages is among the largest schema-driven lifts we observe. The deployment requires creator (Person or Organization), datePublished, license, distribution (with contentUrl to the actual data file or table), and ideally measurementTechnique describing how the data was gathered. The marginal effort is meaningful; the visibility return on original research markup is consistently high.

Several schema types that get heavily promoted in SEO content produce minimal citation lift in our observations. BreadcrumbList helps SERP display but doesn't move citation. Review and AggregateRating help product rich snippets but rarely surface in AI citation. WebPage is largely redundant when more specific types (Article, ProductPage, etc.) are deployed.

Avoid schema sprawl. Sites that deploy 8–12 different schema types per page (often via CMS plugins that 'maximise rich snippets') do not outperform sites that deploy 3–5 well-chosen schema types correctly. The discipline is to pick the schema types that match your content shape and deploy them with full property coverage, not to layer every possible type for theoretical coverage.