Does schema markup help you get cited by AI? No, not directly, and the evidence on this is stronger than the industry admits. Ahrefs tracked 1,885 pages that added JSON-LD schema and found AI citation changes close enough to zero to be classified as random noise. A separate Search Atlas study found no correlation between schema coverage and citation rates across ChatGPT, Gemini and Perplexity. Schema is still worth implementing. It is infrastructure, not a citation lever, and the difference matters when you are deciding where to spend.
What the studies found
Three pieces of evidence matter here, and one widely cited study turns out not to be about schema at all.
Ahrefs, May 2026. Researchers tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matched them against 4,000 control pages with similar citation levels, and measured the 30 days before and after using difference-in-differences analysis plus three further statistical tests. Citation changes in AI Mode and ChatGPT were indistinguishable from noise. Google AI Overviews showed a statistically significant decline, although the absolute effect was small.
Search Atlas. Found no correlation between schema coverage and citation rates across OpenAI, Gemini and Perplexity. Sites with comprehensive schema performed the same as sites with almost none.
The Princeton and Georgia Tech study from 2024. This is the one you will see cited in most GEO articles as proof that schema drives citations. It did not test schema. It tested content strategies, specifically the effect of adding citations, statistics and quotations. The finding is real and useful. It is not about structured data, and anyone citing it as schema evidence has not read it.
As of early 2026 there are no peer-reviewed studies on schema markup and LLM visibility at all. An industry making confident claims is doing so on an empty evidence base.
Why doesn’t schema work the way people assume?
The mechanism explains the result.
Testing by SearchVIU found that major AI systems extract only visible HTML content during retrieval, and ignore JSON-LD. Your schema block sits in the page source describing your content accurately, and the model reads the text a human would read instead.
That makes sense when you think about what a language model does. It processes the rendered content. A hidden metadata layer written for search engine parsers is not what it is looking at.
This single fact explains most of the confusion in this topic. Schema was built to help search engines index and display pages. Language models retrieve and summarize text. Those are different jobs.
What do the platforms themselves say?
Worth separating confirmed statements from assumption.
Microsoft is the only clear yes. In March 2025 Fabrice Canel confirmed that schema markup helps Bing’s LLMs understand content for Copilot. That is a first-party statement and it is specific.
Google has been vague, saying only that structured data offers an advantage in AI-generated search experiences. That is not nothing, but it is not a mechanism either.
OpenAI, Anthropic and Perplexity have said nothing publicly about how they handle schema.
So the confirmed benefit is real and it is narrow: Copilot and Bing. Everything else is extrapolation.
The caveats that cut the other way
Fairness requires stating the limits of the sceptical evidence too.
The Ahrefs study only examined pages that were already being cited, with a hundred or more citations each. It cannot tell you whether schema helps an uncited page get discovered in the first place, which is the situation most businesses are actually in.
It also pooled every schema type together, so type-specific effects were not measured. Organization schema and Recipe schema do very different jobs and the study treated them as one variable.
And schema-using sites tend to be better maintained, with stronger content and more authority. Correlation cuts both ways here.
The honest position is that schema does not appear to be a citation lever, and that nobody has yet proven it is useless for discovery.
So why implement it at all?
Because it does several other things, all of which are worth having.
It disambiguates your entity. Organization schema with a sameAs array is how you tell every machine reading your site that the various profiles mentioning your business are the same business. Entity clarity is one of the things that genuinely does influence whether a model can recommend you by name.
It feeds Bing and Copilot, the one confirmed case.
It enables rich results, which affect click-through rate in ordinary search and still matter for the roughly 87% of searches that show no AI Overview.
It feeds the Knowledge Graph and voice assistants.
It is cheap and permanent. A correct Organization block is an afternoon’s work and then it is done.
None of that is a growth hack. All of it is infrastructure, and infrastructure is worth having.
The types actually worth your time
Ordered by what they do, rather than by any claimed citation effect.
Type | What it actually does | Evidence level | Priority |
|---|---|---|---|
Organization | Establishes you as a single identifiable entity; sameAs links your profiles together | Strong reasoning, no direct citation study | Do it first |
LocalBusiness | Feeds local packs, Maps and local AI recommendations | Well established for local search | Essential if you have a location |
Person | Ties content to a named author with credentials; supports E-E-A-T | Indirect but well supported | High |
Article / BlogPosting | Author, publisher, publish and modified dates | Standard practice | High |
Product | Price, availability, reviews; feeds shopping surfaces and AI shopping agents | Strong for commerce | Essential for ecommerce |
Review / AggregateRating | Surfaces review counts and scores as machine-readable corroboration | Strong for rich results | High if you have reviews |
BreadcrumbList | Communicates site structure and where a page sits | Standard practice | Medium, and trivially easy |
Service | Describes what you sell and where you sell it | Weak evidence, sensible | Medium for service businesses |
FAQPage | Marks up question and answer pairs — see the section below | Rich results deprecated May 2026; no citation evidence | Low, but nearly free |
About FAQ schema specifically
This needs updating, because Google deprecated FAQ rich results in May 2026.
The markup no longer produces the expandable questions in search results for most sites, which was its main visible benefit. Combined with the citation evidence above, the case for FAQPage markup is weaker than it was a year ago.
But the visible FAQ content still matters a great deal, and for a reason that has nothing to do with schema. AI systems extract visible text. A clearly written question with a self-contained 40 to 60 word answer sitting in your page copy is exactly the format a model can lift. The question section works. The markup around it is the part that has lost its value.
So keep writing FAQ sections. Keep adding the markup, because it is nearly free and Bing reads it. Stop expecting either to produce citations on its own.
What actually influences citations
If schema is not the lever, what is?
The evidence points at content and corroboration rather than markup. Answer the question in your opening sentences. Include specific numbers, dates and named entities. Make claims that survive being quoted out of context. Earn mentions on sites you do not own. Build review volume. Those are the things the Princeton study actually tested, and the things that keep showing up in every analysis of what cited pages have in common.
Schema supports all of that by making your identity unambiguous. It does not substitute for any of it.
Want to know what is actually holding you back?
We will run twenty buyer prompts across ChatGPT, Gemini, Perplexity and Copilot, check whether you appear, and tell you which of your signals is the weak one. Often it is not your markup.
Frequently Asked Questions
Does schema markup help with AI citations?
Not directly, based on current evidence. Ahrefs tracked 1,885 pages adding JSON-LD against 4,000 matched controls and found citation changes in AI Mode and ChatGPT indistinguishable from random noise. Search Atlas found no correlation between schema coverage and citation rates. Schema remains useful as infrastructure, but it is not a citation lever.
Why don’t AI systems read my schema markup?
Because they extract visible HTML content during retrieval and largely ignore JSON-LD. Testing by SearchVIU found this pattern across major AI systems. Schema was designed to help search engines index and display pages; language models retrieve and summarize rendered text. They are different jobs.
Is schema markup still worth implementing in 2026?
Yes, as infrastructure. It disambiguates your business as a single entity through Organization and sameAs, it is confirmed to help Bing and Copilot, it enables rich results that still affect click-through on the majority of searches showing no AI Overview, and it feeds the Knowledge Graph. Just do not expect citations from it.
Has FAQ schema been deprecated?
Google deprecated FAQ rich results in May 2026, so the markup no longer produces expandable questions in search results for most sites. The markup is still read by Bing and costs almost nothing to keep. The visible FAQ text remains genuinely valuable, because AI systems extract visible content — but that benefit comes from the copy, not the markup.
Which schema types should I prioritise?
Organization first, with a complete sameAs array. Then LocalBusiness if you have a location, Person on your authors, Article on your content, and Product plus Review if you sell online. BreadcrumbList is trivially easy and worth adding. FAQPage is low priority now but nearly free.
What actually makes AI assistants cite my site?
Answering the question in your opening sentences, including specific numbers and named entities, writing claims that survive being quoted out of context, earning mentions on sites you do not own, and building review volume. Those are the variables with evidence behind them. Markup supports them rather than replacing them.