• 1 min read

JSON-LD for AI search

Audit and extend Schema.org JSON-LD so retrieval pipelines can classify your pages - which types matter, and why incomplete markup is the usual fault.

Quick answer: Complete Schema.org JSON-LD matching visible page content is a sensible place to start, because search indexes already consume it.

  • Machine-Readable Delivery
  • AI Search Visibility
  • Structured Data
Jump to section
  1. Why JSON-LD Comes First
  2. JSON-LD for AI Search: Common Questions

Quick answer

What JSON-LD should a site publish to be cited by AI search?

High confidenceVerified 1 Sept 2026
Publish Schema.org JSON-LD that describes the entities a page is genuinely about - Organisation, Article with a named author, Product with price and availability, FAQPage, BreadcrumbList - and keep every property consistent with the visible page text.

Sources

Why It Comes First

Why JSON-LD Comes First

Of the machine-readable layers now competing for attention, structured data is the only one with an established consumer. Search engines have read Schema.org markup since 2011, and because AI retrieval largely runs against those same indexes, the markup that helps a crawler understand a page is what reaches an assistant deciding whether to quote it. The newer formats are bets on future adoption. This one is already being read.

It also sits first in the wider sequence of work that keeps a site visible to AI assistants, ahead of the newer protocols - though it is worth confirming which assistants actually fetch your pages before deciding how much of this to do at once.

The practical gap on most sites is not absence but incompleteness. An Organisation block exists because a theme generated one; Article markup omits the author; Product markup carries a name but no price or availability. Partial markup describes a page badly, which is worse than describing it not at all, because a retrieval pipeline treats what it finds as authoritative.

The Consistency Rule

Structured data must describe what a reader actually sees. Markup asserting a price the page does not show, a rating nobody left, or an author who does not appear is treated as misrepresentation and can cost rich-result eligibility. The rule is simple enough to audit: every property in the JSON-LD should be traceable to something visible on the rendered page. This matters more once the consumer is an assistant, because a model that quotes a mismatched figure attributes it to your brand.

This is also where a replatform tends to go wrong. Markup is usually generated by templates, so a front-end rebuild silently drops it - the same class of regression that makes preserving search rankings through a replatform a discipline of its own rather than a checklist item.

How to audit and extend your JSON-LD

Typically one to three days for a mid-sized site, depending on template countIntermediate

A working sequence for taking structured data from partial to complete, in the order that surfaces problems earliest.

Tools Required

  • A site crawler that extracts JSON-LD
  • A structured-data validator
  • Access to the page templates

Steps

  1. 1

    Inventory what already exists

    Crawl the site and extract every JSON-LD block by template rather than by page. Most sites have a handful of templates generating everything, so the inventory is far smaller than the page count suggests.

  2. 2

    Map each template to the types it should carry

    An article template needs Article with author and dates; a product template needs Product with offers; a service page needs Organisation and BreadcrumbList. Decide the target per template before editing anything.

  3. 3

    Fill the required properties, not just the optional ones

    Incomplete types are the common failure. Article without author, Product without offers and FAQPage without answers all parse successfully while conveying almost nothing.

  4. 4

    Check every property against the visible page

    Remove or correct anything the reader cannot see. Markup that asserts more than the page shows is a policy breach, not a shortcut.

  5. 5

    Validate and re-crawl

    Run the markup through a structured-data validator, then re-crawl to confirm the fix reached every page the template serves rather than only the one you tested.

Key Takeaways

What Actually Moves the Needle

  • Structured data is the only machine-readable layer with an established consumerCritical

    Search engines have read Schema.org since 2011, and AI retrieval largely runs on those indexes. The newer protocols are bets on adoption; this one is already read.

  • Incomplete markup is the common failure, not missing markupCritical

    Article without an author, Product without offers, FAQPage without answers - each parses cleanly while telling a retrieval pipeline almost nothing about the page.

  • Every property must be traceable to something the reader can seeImportant

    Markup asserting a price, rating or author absent from the page risks rich-result eligibility, and an assistant that quotes the mismatch attributes it to your brand.

  • Audit by template, not by pageImportant

    A small number of templates generate the whole site, so fixing a template corrects thousands of pages - and testing a single page proves nothing about the rest.

Complete, honest structured data on every template is among the more dependable places to start, because it is the only layer already being consumed at scale.

JSON-LD for AI Search: Common Questions

Does JSON-LD actually affect whether an AI assistant cites us?
Indirectly, and only ever as one input among many. Assistants retrieve against conventional search indexes, so markup that helps those indexes classify a page correctly improves the chance it surfaces as a candidate at all. Structured data does not force a citation, but a page a retrieval pipeline cannot classify rarely becomes one.
Which Schema.org types are worth the effort?
Start with the ones describing what your organisation and pages genuinely are: Organisation, BreadcrumbList, and then the type matching each template - Article with a named author for editorial pages, Product with offers for commerce, FAQPage where real questions and answers appear. Extending beyond the types that match visible content adds noise rather than signal.
Is Microdata or RDFa acceptable instead of JSON-LD?
They are still parsed, but JSON-LD is the format to standardise on. It sits in a single script block rather than being interleaved through the markup, which means a template change is far less likely to break it and an audit can extract it cleanly.
What happens if the markup does not match the page?
It is treated as misrepresentation and can cost rich-result eligibility. The more serious risk with AI retrieval is that a model quotes the asserted figure and attributes it to your brand, so an inaccurate property becomes an inaccurate public statement rather than a silently ignored one.
How often should structured data be re-audited?
Whenever templates change, since markup is generated by templates and a front-end rebuild is the most common way it silently disappears. Beyond that, an annual pass tends to catch drift such as authors who have left or products no longer sold.

Want a structured-data audit?

Written by Cameron Young, CEO

  • 1 min read

Related reading