Schema for AI search — what still matters after Google's 2026 guidance

Google’s 2026 guidance is that structured data isn’t strictly required to appear in AI features (Our Name Is Mud). That has been reported as “schema doesn’t matter for AI” — which is not what the guidance says, and not what the 2026 evidence shows. Schema markup was officially confirmed as “AI infrastructure” by Google, Microsoft and OpenAI in March 2025, and multiple 2026 analyses find it materially enriches the knowledge graphs LLMs draw from (Blck Alpaca, Whitehat SEO).

The schema types that actually earn citations

Article — the workhorse. Include headline, author (as a linked Person), datePublished, dateModified, image, mainEntityOfPage and keywords. Strapi’s guide argues content pipelines should store JSON-LD alongside the content itself, populated at write time, and served through the same API that renders the page (Strapi).

FAQPage — very high citation return relative to effort. Blck Alpaca reports FAQPage as one of the highest-leverage AI-search schemas of 2026 (Blck Alpaca).

HowTo — for procedural content. Especially strong for Perplexity and Copilot, which favour explicit step lists.

Organization with sameAs. The sameAs array — pointing to Wikipedia, Wikidata, Crunchbase, LinkedIn, GitHub — is one of the strongest entity-disambiguation signals for AI systems.

Person on author bios. Link authors to LinkedIn, personal sites and past work. E-E-A-T is doing more for AI citations than for classical rankings (Stackmatix).

Ship it in JSON-LD, not Microdata

Use JSON-LD in a <script type="application/ld+json"> block. It is what Google prefers, what most 2026 tooling generates, and it decouples structured data from your HTML layout so you can change one without breaking the other.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "Answer-first writing — the single highest-leverage GEO change",
  "author": { "@type": "Organization", "name": "GEOhandbook" },
  "datePublished": "2026-09-15",
  "mainEntityOfPage": "https://geohandbook.io/articles/answer-first-writing/"
}
</script>

Validate in CI

If schema is worth adding, it is worth failing your build when it breaks. Strapi’s guide walks through a GitHub Actions setup that extracts JSON-LD from rendered HTML, asserts required properties exist, and fails the build on error — with a stated pipeline overhead of under 30 seconds (Strapi).

Common mistakes

Schema that doesn’t match the visible page. Google will demote or ignore it; AI systems will notice too.
FAQPage with fake questions. Only mark up genuine FAQ content.
Person without sameAs. A bare author name is a low-signal entity. Link it out.
Multiple @id values for the same entity across pages. Pick one canonical ID and reuse.

Checklist

[ ] Article on posts with author, datePublished, dateModified.
[ ] FAQPage on FAQ sections.
[ ] Organization on the homepage with sameAs.
[ ] Person on author bios with sameAs.
[ ] JSON-LD validates on the Rich Results Test and Schema.org validator.
[ ] Validation runs in CI on every pull request.