Schema Markup for AI: What the Evidence Actually Shows

Search for "schema for AI indexing" and you will find a market: generators, plugins, agencies, and audit tools — ours included, until this month — telling you that JSON-LD is how you get cited by ChatGPT. This post is the counter-position, and it is not a hot take. It is what the operator of the largest AI search surface says in writing, plus the only controlled experiment anyone has run.

What Google says, verbatim

From Google's guidance on AI features in Search, updated 2026-07-10:

"Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add."

And, on eligibility for AI Overviews and AI Mode (developers.google.com/search/docs/appearance/ai-features):

"To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements."

Google is an interested party — its statements are assertions, not measurements. But they are the only first-party statements that exist, and they are unambiguous: no AI-specific schema type exists, and markup is not a requirement.

What the one controlled test found

Ahrefs published the only matched-control study we know of on 2026-05-11. Design: 1,885 pages that added schema, each matched to three control pages. Result, in their words: "Adding schema produced no major uplift in citations on any platform." AI Overview citations fell 4.6% on treated pages — statistically significant. AI Mode moved +2.4% and ChatGPT +2.2%, neither significant.

One boundary condition, which Ahrefs disclosed and we repeat: every page in the dataset already had 100+ AI Overview citations before any schema was added. So the study tests whether schema helps a page that is already cited get cited more. Whether it helps an uncited page enter the consideration set is untested — not refuted. If you are a mid-market brand at zero citations, nobody has run your experiment. That is an honest gap, and it is not the same thing as evidence in your favour.

FAQ schema specifically

The single most-sold recommendation in GEO was "add FAQPage schema." Google deprecated FAQ rich results on 2026-05-07 and removed the FAQPage documentation on 2026-06-15 — the old documentation URL now redirects to a changelog entry. Search Console dropped API support in August 2026. The markup no longer produces anything in Google Search. Google says leaving existing markup in place causes no harm, so there is no reason to rip it out. There is also no reason to add it.

Why the industry sells it anyway

Four reasons, none of them evidence.

• It is easy to generate and easy to check. A tool can emit JSON-LD in a second and detect it in a second. That makes it a perfect audit line item, which is a fact about audit tools, not about AI.
• It borrows from a real SEO mechanism. Schema did earn rich results in classic Google Search for years. The category transferred the intuition to AI answers without checking whether the pipeline was the same. It is not: ChatGPT, Perplexity and Claude crawlers read raw HTML text and no engine documents parsing JSON-LD for facts.
• The contamination is structural. Most GEO frameworks are published by companies that sell the fix the framework recommends. That includes us. It is why we treat our own on-page recommendations with the harshest scrutiny.
• The story is intuitive. "Without schema, AI has to guess, and when AI guesses it hallucinates" reads well. No mechanism test has found any engine retrieving a fact that exists only in JSON-LD, so the causal chain has no first link.

What schema is still good for

Three things, all in Google Search, none of them AI.

• Rich results that still exist. Product snippets and merchant listings (Product with Offer), review snippets where eligible, breadcrumbs, article metadata. These render in classic results and are worth having for that reason.
• Disambiguation. If your brand name is also a common word, Organization schema with sameAs links tells Google which one you are. This is the honest use of Organization markup.
• Machine-readable dates and authorship for Google's article surfaces — provided they match the visible page, which brings us to the one thing that matters for AI.

The one defensible AI use: parity

Google's guidelines have always required structured data to reflect what a user sees on the page. That rule turns out to be the only schema-related check with a plausible AI mechanism, because it catches two failure modes that do reach AI answers:

1. A fact that lives only in the schema. No AI crawler will ever read it. If your aggregate rating, founding year or price is in the JSON-LD and nowhere in the visible text, it is invisible to every engine except Google's — and Google says it does not need it.
2. A fact that disagrees with the page. This is a published contradiction. An engine that fetches your page may read one number in the text; a tool that parses your schema reads another. You cannot control which one wins.

A worked example

A pricing page carries this Product schema:

{
  "@type": "Product",
  "name": "Growth Plan",
  "aggregateRating": { "@type": "AggregateRating", "ratingValue": "4.8", "reviewCount": "120" },
  "offers": { "@type": "Offer", "price": "99.00", "priceCurrency": "USD", "priceValidUntil": "2026-12-31" }
}

Extract the facts and check each against the raw HTML with JavaScript disabled:

Fact in schema	In visible text?	Verdict
Growth Plan, $99.00 USD	Yes — pricing table reads "$99/mo"	Parity
Rating 4.8 from 120 reviews	No — appears nowhere on the page	Invisible to AI; also a guideline violation
Price valid until 2026-12-31	No — page says nothing about validity	Invisible to AI

Two of three facts exist only in the markup. Neither will ever appear in an AI answer, and the rating is the kind of claim Google's guidelines say must be visible. The fix is not more schema. It is either putting the facts on the page or removing them from the markup. Now imagine the table read "$89/mo" while the schema said 99.00 — that is the contradiction that shows up in an audit as "ChatGPT gets our pricing wrong."

What we are changing

Our Schema Markup Generator is being repositioned. It used to be described as generating markup "optimized for AI citation." It now does two things: generate valid JSON-LD for the Google rich results that still exist, and check that every fact in it also appears in your visible text. The audit engine's schema rule is being rewritten the same way — from "add FAQ schema" to "these facts are in your markup but not on your page." We had the most-emitted and least-evidenced recommendation in the product. We would rather say so than keep shipping it.

What to do instead

• Keep Product and Organization schema for Google's remaining rich results and for disambiguation. Skip FAQPage and HowTo; both are retired.
• Run the parity check on every page that carries JSON-LD. Put invisible facts on the page or delete them from the markup.
• Spend the time you save on access: fetch your pages with JavaScript disabled, resolve robots.txt per AI retrieval bot, and request a page as each bot to confirm your CDN is not returning 403. Those are the levers with a documented mechanism — see how LLMs prioritise information about brands.

Run a free audit and we will list the facts that are in your schema but not on your page.