Will this guarantee my page is cited by ChatGPT or Perplexity?

No. Schema markup is necessary but not sufficient. It removes ambiguity at the indexing stage. Citation also requires answer-shaped content, freshness, and entity authority. See the AI Search Refusal Patterns reference for the full picture.

I read that chatbots ignore JSON-LD on direct fetch. Is this still worth it?

Yes. The Searchviu 2025 test showed that direct fetch by some chatbots ignores JSON-LD, but training and indexing phases (which feed Google AI Overviews, Bing Copilot, and most retrieval pipelines) do consume it. The signals also shape entity-coherence priors that affect citation even when JSON-LD is not parsed live.

How many sameAs links should I include?

For an Organization: at least three high-quality links — ideally Wikidata or Wikipedia plus two professional networks (LinkedIn, GitHub, Crunchbase). For a Person: at least one verifiable identifier (ORCID, Google Scholar, LinkedIn). Quality outweighs quantity; padding with low-trust profiles can hurt.

Do I need a separate JSON-LD block per entity?

No — use a single @graph per page so engines can resolve @id cross-references in one parse. Fragmented blocks force engines to deduplicate and risk losing a node entirely.

How often should I update my graph?

Update dateModified on Article nodes whenever the visible content changes. Re-emit Organization and Person graphs only when their facts change (logo, role, credentials). Treat the canonical topic graph (Pattern 4) like product data: versioned, reviewed, monitored.

What about Action and PotentialAction?

Out of scope here. These patterns require a behavioural integration (e.g. SearchAction wired to a working endpoint) and rarely move the needle on AI-search citation. Add them only if you have a real action surface to expose.

Knowledge Graph Markup for AI Search: A schema.org Pattern Specification

Knowledge graph markup for AI search uses JSON-LD on schema.org to declare entities and link them through @id, sameAs, about, mentions, and citation properties. The goal is not rich snippets — it is to give generative engines an unambiguous machine-readable identity layer at the indexing stage so they can ground and cite your content with confidence.

TL;DR

Five composable patterns (Organization Hub, Author Identity, Article Citation Chain, DefinedTerm Glossary, WebPage about/mentions) cover most knowledge-graph markup needs for AI search.
Use stable canonical @id URIs as primary keys. Cross-link them with sameAs, about, mentions, and citation.
Schema markup does not guarantee citation. It removes ambiguity at indexing time, which is necessary but not sufficient.

Status of this specification

This is an opinionated implementation specification, not a tutorial. It assumes familiarity with JSON-LD and schema.org. Where schema.org defines multiple ways to express a relationship, this spec selects one. Conformance means matching the patterns and constraints below.

Scope and non-goals

In scope: entity declaration and linking patterns that are demonstrably useful at the indexing and retrieval stages of generative engine pipelines (Google AI Overviews, Bing/Copilot, ChatGPT Search, Perplexity, Gemini, Claude search modes).

Out of scope: rich-result-only properties (e.g. Recipe.cookTime), pre-AI Google features Google has retired, and any Action/PotentialAction patterns that require behavioural integration.

Honest caveat: an October 2025 Searchviu test found that during direct-fetch (a chatbot pulling a single URL on demand), most chatbots extract only visible HTML and ignore JSON-LD. Schema markup still matters during the training and indexing phases that feed retrieval, especially for engines that use Google's index (AI Overviews, Bing Copilot). Treat this spec as a long-term identity layer, not a same-day visibility hack.

Conventions

All examples use JSON-LD in