Schema Markup for AI Search: A Technical Guide

That gap between "not a ranking factor" and "the cited pages tend to have it" defines schema work in 2026. Answer engines assemble responses from search indexes that structured data has fed for a decade. Skipping it does not make you invisible. It makes you harder to disambiguate, and engines resolve ambiguity by citing someone else.

How AI engines use structured data

Three pipelines matter. Google AI Overviews sit on top of the regular Google index, where structured data has powered rich results and the Knowledge Graph since 2012. Bing feeds both Copilot and ChatGPT's live search mode, and Bing's engineers have said in public that schema helps their language models understand page content. Training crawlers such as GPTBot and ClaudeBot mostly reduce pages to text, so markup matters less there.

The practical consequence: schema works on the retrieval side, before any engine decides whom to quote. It determines whether the index understands who you are and what the page answers. Entity clarity compounds too. A consistent Organization block with a sameAs array ties your site, your LinkedIn page and your directory listings into one entity instead of four guesses.

The five schema types that matter for AI search

• Organization (or ProfessionalService) on the homepage, with name, url, logo and a sameAs array pointing to every real profile you control. This block anchors entity disambiguation.
• Article or BlogPosting on every post, with headline, datePublished, dateModified and an author that references a real person with a real profile URL. Authorship is a signal engines can verify.
• FAQPage on pages with visible question and answer pairs. Google cut the FAQ rich result for most sites in 2023, but the machine-readable Q&A format maps one to one onto how answer engines retrieve passages.
• BreadcrumbList, so engines know where a page sits in your hierarchy and attribute an answer to the right section of your site.
• Service or Product with offers, price and availability where the page shows them. Commercial queries in AI engines pull these fields into comparisons.

How to implement JSON-LD without it rotting

Use JSON-LD in a single script tag, the format Google recommends, and combine types on one page with @graph instead of stacking separate blocks. Then generate it from the same data source that renders the visible content. On our site, the blog's JSON-LD reads from the same TypeScript files that render each page, so a title change updates both at once. Hardcoded schema drifts, and drift is what validators flag and engines discard.

On React or Next.js, render the script tag on the server so crawlers get it in the initial HTML. Client-side injection reaches Googlebot after a render queue, and reaches lighter AI crawlers late or never.

Validate before you ship

Run every template through Google's Rich Results Test and the schema.org validator before deploying, then watch the enhancement reports in Search Console for regressions. The errors worth checking twice: dates outside ISO 8601, relative instead of absolute URLs, and enumeration values typed from memory. One malformed field fails the whole block, and nothing on the page warns you.

The implementation checklist

• Pick the pages that earn markup first: homepage, blog posts, service pages, FAQ.
• Generate JSON-LD from your content source instead of pasting static blocks.
• Combine types per page with @graph in one script tag, rendered server-side.
• Reference real authors with profile URLs in every Article block.
• Validate with the Rich Results Test and the schema.org validator.
• Recheck the Search Console enhancement reports two weeks after each deploy.

Schema is one layer of the machine-readable stack

AI crawlers also read llms.txt, a plain-text index of your site written for language models. Adoption is early: 8.7% of the top 1,000 sites had one by June 2026, and the count across the web grew 8.8x in twelve months to about 36,000 sites. On mindworks.cl we ship all three layers: JSON-LD on every post and FAQ, an llms.txt, and robots rules that name each AI bot instead of treating them as one crawler.

Does Google use schema markup for AI Overviews?

Google has not named structured data as an AI Overviews ranking factor. The overviews draw from the same index and the same understanding layer that schema feeds, and the 20% to 30% appearance gap for pages with valid markup shows up across independent 2025 benchmarks. Treat it as index infrastructure rather than an AI switch.

Do LLMs read JSON-LD directly?

It depends on the pipeline. Crawlers collecting training data mostly reduce pages to plain text. Engines that search before answering, like Copilot, ChatGPT's search mode and AI Overviews, retrieve through indexes built with structured data, and that is where your markup does its work. Bing has confirmed schema helps its models interpret pages.

Is JSON-LD better than microdata for AI search?

Yes. Google recommends JSON-LD, it lives in one block instead of attributes scattered through your HTML, and you can generate it from a data source and keep it in sync with the visible content. Migrating old microdata costs about half a day and removes a maintenance trap.

Which pages should you mark up first?

The homepage with Organization, then the blog with Article blocks and author references, then any page with visible Q&A pairs as FAQPage. That order front-loads entity clarity and authorship, the two signals answer engines lean on when deciding whether a source is worth quoting.

Where schema fits in your AEO work

Structured data is the cheap layer: a few templates, one validation pass, and it holds while you do the recurring work of publishing answers worth citing. Our answer engine optimization guide covers that content side, and our comparison of AEO platforms covers measuring the result. If you want to know how your site reads to the machines today, our free growth audit includes a structured data and llms.txt check, with no retainer attached.