Schema Markup for AI Search

What structured data is, and what it is being sold as

Structured data is machine-readable metadata added to a web page describing what the page is about — a product, a person, an organization, an event — using a shared vocabulary. On the web that vocabulary is almost always schema.org, launched on 2 June 2011 by Google, Microsoft and Yahoo, and the dominant serialization is JSON-LD (JavaScript Object Notation for Linked Data), a block of JSON placed inside a script tag on the page.

Its historical payoff was concrete and visible. Correct markup made a page eligible for a rich result — stars, prices, images, an expandable list of questions — in Google's search results, and that eligibility was reportable in Search Console. Everyone who worked in search between 2015 and 2023 learned the same lesson: markup produces a visible search feature, the feature produces clicks, the work pays for itself.

The question this page answers is a different one, and it is the question the generative engine optimization market — GEO, the practice of trying to influence what AI answer systems cite — has been selling against since 2024. Does the same markup change whether a page is retrieved, quoted or cited by an AI answer system? The answer is unusually well evidenced for this subject. Google states in two separate documents that no special schema.org markup is required for its AI features, and the only controlled experiment anyone has published found no citation uplift on any platform. Structured data still earns its keep. It does not earn it there.

Three places markup could act, and only one is documented

Nearly every bad claim on this subject comes from collapsing three distinct mechanisms into one. They are worth holding apart, because they have very different evidence behind them.

As a rich-result trigger. Google validates markup against a per-feature specification and a page that passes becomes eligible for a specific search appearance. This path is real, it is measurable, and it is shrinking — FAQ and HowTo rich results are both gone. It has nothing to do with AI answers.

As an input to entity understanding. Markup that names an entity and links it outward with sameAs hands a reconciliation system a candidate identifier. The mechanism is sound, and it is the argument practitioners working on entity grounding actually make. It is also unconfirmed: Google has never said markup is required for Knowledge Graph reconciliation, and it plainly is not, since most reconciled entities carry none. Strongest theory, weakest direct evidence.

As text in the retrieval pipeline. This is where the AI-citation claim lives and it is the weakest link of the three. For a page to be cited, some passage of it has to be retrieved and placed in a model's context window. On Google's AI features the grounding corpus is Google's own web index, which means the markup was parsed at indexing time and whatever it contributed has already been absorbed into how the page is understood and ranked. Adding markup does not hand the generative layer a new document. It is the same index entry.

The asymmetry worth carrying away: structured data is a claim about a page, and retrieval operates on the page's text. A system that already has the text gains little from a restatement of it. The only place markup has clear leverage is where a machine needs a fact the prose does not state unambiguously — a price, a date, an identifier, a canonical URL.

Google has answered this in writing, twice, in both directions

The AI features and your website page, carrying a last-updated stamp of 10 December 2025, says:

There's also no special schema.org structured data that you need to add.

The guide to optimizing for generative AI features on Google Search, published 15 May 2026 and last updated 10 July 2026, repeats it under a heading Google itself titled Mythbusting generative AI search: what you don't need to do — a section written specifically to list the things publishers are being told to do that Google says are unnecessary. Its entry on this subject reads:

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

And immediately after it, the half practitioners skip:

However, it's a good idea to continue using it as part of your overall SEO strategy, as it helps with being eligible for rich results on Google Search.

That is Google's complete position and it is unambiguous in both directions. Structured data is not an AI-visibility lever, and it is not obsolete. Confusing not required with does nothing is the error that leads teams to rip out working Product, Review and Event markup because they read that AI had killed schema.

The silence elsewhere is just as informative. OpenAI documents four crawlers and says nothing about structured data. Anthropic's crawler page, updated 7 April 2026, says nothing. Perplexity's bot documentation says nothing. Microsoft's Bing Webmaster Guidelines, rewritten in February 2026 to cover Copilot, discuss structured data in general SEO terms without claiming it helps Copilot cite a page. Four operators have published publisher-facing crawler documentation in the last eighteen months. If markup were a lever they controlled, that is where they would say so.

The one experiment, and the trap inside it

On 11 May 2026 Ahrefs published what appears to be the only controlled test of the question. Louise Linehan and Xibeijia Guan took 1,885 pages that added JSON-LD between August 2025 and March 2026 and compared them against roughly 4,000 matched control pages that did not, measuring citations for thirty days either side using difference-in-differences, t-tests and event studies.

The result: AI Overviews citations changed by −4.6%, AI Mode by +2.4% and ChatGPT by +2.2%. The authors' own summary is that adding schema produced no major uplift in citations on any platform. Noise in three directions is what a null looks like.

The same publication also contains the number that half the industry is quoting. Across roughly six million URLs in its correlational arm, cited pages were about three times as likely to carry JSON-LD as uncited ones, and 53% of AI-cited pages had schema of some kind. Both halves are in the same post. Quoting the 3x figure while omitting the experiment is the single most common error on this subject in 2026, and it is not a subtle one — the study was designed to show precisely that the correlation does not survive a controlled comparison. Big, well-maintained sites carry both schema and citations for the same underlying reason: they are big, well-maintained sites.

The limits the authors state should travel with the numbers. The sample was restricted to pages already receiving a hundred or more AI Overview citations, so it measures the effect on pages already visible to AI systems rather than on obscure ones. Schema additions often coincide with other page changes. Schema types were pooled rather than tested individually. The window was thirty days. Only JSON-LD was tested. One well-designed null is not a settled law, and anyone reading this page should hold it as one good study, not replicated rather than as proof.

Where the 112% figure came from

The most-repeated number in this market is a claimed 112% increase in AI citations from FAQ schema. It is worth tracing, because the tracing is the whole lesson.

The figure originates in a study published by the agency TechCognate on how Gemini selects citations. It is correlational — it observes that cited pages tend to carry FAQ markup — and it is an unattributed agency publication, which ranks below platform documentation, platform blogs, named studies with published methodology, and named journalists. It appears across dozens of GEO posts, usually without attribution and almost always without the word correlational. It is contradicted by the only controlled test on the question.

It spread because it is large, quotable, directional, and it arrived exactly when practitioners wanted permission to keep doing something familiar. That is the standard life cycle of a claim in this field: a plausible correlation, an agency post, a round number, and eighteen months of citation-free repetition. What would settle it: a matched-pair test that adds only FAQPage markup to otherwise unchanged pages and measures citation change. Nobody has published one, which is itself worth reporting.

What happens to JSON-LD when an assistant fetches your page

There is a mechanical argument here that almost nobody acts on, and it cuts directly against the standard advice.

When an assistant fetches a live URL because a user just asked about it — ChatGPT-User, Claude-User, Perplexity-User — the page is converted from HTML to text before it enters a context window. Standard HTML-to-text conversion strips script tags. JSON-LD is a script tag. Markup delivered as JSON-LD is therefore plausibly invisible to exactly the systems people deploy it for, while Microdata and RDFa, which live in the visible markup, survive conversion better.

No operator publishes its extraction pipeline, so treat that as a mechanism argument rather than a documented fact — a reason to expect the null result Ahrefs found, not independent proof of it. But it is cheap to test and nobody has published a test. Put a distinctive fact in a page's JSON-LD, state it nowhere in the visible text, ask each assistant about it, and see whether the fact comes back. The absence of any published instance of that experiment, on a question this commercially significant, says something about how much of the GEO market is built on hypotheses that were never checked.

What structured data is still worth deploying for

None of the above is an argument for removing markup, and the advice to strip it is worse than the advice to add it, because removal has a real cost — template changes, regression risk, lost rich results — against a benefit of zero.

• Rich results that still exist still pay. Product, Review, Recipe, Event, JobPosting, Breadcrumb, Video and LocalBusiness markup all drive live search appearances. The deprecation of two feature types is not a verdict on the format.
• Unused markup is harmless. Google's standing position, given when it deprecated FAQ and HowTo rich results, is that while you can drop the structured data there is no need to proactively remove it, and that structured data which is not being used does not cause problems for Search.
• Do not add types with no live consumer. FAQPage has produced no Google search appearance since 7 May 2026 and HowTo none since 14 September 2023. Adding either now buys a maintenance burden and no surface.
• Markup must match the visible page. This is the one structured data rule with an enforcement mechanism behind it — mismatched markup is a spam-policy problem, not a missed opportunity. It is also why marking up content you do not actually have is bad advice regardless of what it might theoretically do for AI.
• Say it in the prose. If a fact matters to a retrieval system, it needs to be in the text a retrieval system reads. Markup is a restatement, not a channel.

The honest position, as of August 2026, is narrow and defensible: the correlation between schema and AI citation is real and large, the causal evidence is a single well-designed null, and the platform that owns the biggest AI answer surface has written down twice that its markup requirement is none. Anyone billing schema implementation as an AI-visibility deliverable is selling something Google has said in writing does not do what they claim.

Frequently asked questions