Ask ten GEO consultants how to get cited by AI and nine will tell you to add schema markup. Ask them for the study that proves it works and the room goes quiet. Structured data occupies a strange position in AI discoverability: universally recommended, rarely tested, and misunderstood in both directions. Some teams treat it as a magic switch. Others, having seen no immediate effect, dismiss it entirely.
Both are wrong, and the reason is that structured data does not do what most people think it does. It does not make AI rank you higher. It makes AI less likely to get you wrong. This article lays out what the major AI companies have actually said, what the evidence does and does not show, and how to use structured data for what it is genuinely good at.
What Structured Data Is, and What It Is For
Structured Data — Machine-readable annotations, most commonly written in JSON-LD using the schema.org vocabulary, that state facts about a page in a form software can parse without interpreting prose: this is an Organization named X, with logo Y, whose official profiles are at Z; this page is an Article by author A published on date B; this product costs C. Search engines and AI systems use it to disambiguate entities and extract facts, not to judge quality.
The distinction matters. When Google introduced its Knowledge Graph in 2012, it described the goal as understanding "things, not strings." Structured data is how a website tells a machine which thing it is talking about. An AI model that has learned about your company from thousands of scattered mentions, some of them conflicting, benefits from a single authoritative statement of the basic facts. That is the job.
What the AI Companies Have Actually Said
The public record is thinner than the advice industry suggests, but it is not empty.
Google has said the most, and said it carefully. In April 2025, Google's John Mueller told a Search Central audience that site owners should keep using supported structured data types because they make content easier for computers to read and index, even as AI features expand. Asked directly in January 2026 whether schema helps large language models, his answer was "yes, no, and it depends," noting that some AI features, such as shopping results, depend heavily on structured data while others do not. Google's documentation names JSON-LD as its only recommended format and describes structured data as helping it understand "the content of the page and to gather information about the web and the world in general."
Microsoft has been the most direct. Fabrice Canel, a principal product manager on Bing, confirmed in March 2025 that Microsoft uses schema markup to help its large language models understand page content, and recommended the IndexNow protocol so generative systems can verify freshness against their training data. Given that Copilot draws on Bing's index, this is the clearest statement from any AI provider that schema is consumed.
OpenAI, Anthropic, and Perplexity have published documentation about their crawlers and how to control them, but none has published guidance saying that schema markup influences what their models cite. Claims that ChatGPT search "values FAQPage schema" circulate widely and trace back to no OpenAI statement.
The honest summary of the public record: structured data is consumed by the systems that build the indexes AI answers draw from, and nobody has shown that it is rewarded by the models that write the answers.
What the Evidence Shows
There is no independent, controlled study isolating schema markup as a cause of increased AI citations. Figures such as "schema increases AI citations by 30%" or "marked-up pages are cited 3.2 times more often" appear on marketing blogs without a disclosed methodology, and a frequently cited "Semrush study of ten million URLs" showing pages with Article and FAQPage schema earning 2.3 times more AI Overview citations cannot be traced to any Semrush publication. Treat all of these as unverified.
What the evidence does support is the underlying problem structured data addresses. The Columbia Journalism Review's Tow Center tested eight AI search tools against 1,600 queries in 2025 and found they answered more than 60% incorrectly, often with fabricated or misattributed citations, and rarely acknowledged uncertainty. In QuadrantX's own category data, large organisations routinely appear as several different entities: a bank, its asset management arm, and its capital markets division scored separately, or a fund provider and its flagship product treated as competitors. The models are not being careless. They are reproducing the inconsistency of what they read.
That is the shape of an entity-consistency problem, and it is the problem structured data was designed to solve. A machine-readable statement of your pricing, your product names, and your corporate identity is not a citation magnet. It is the reference copy that reduces the chance of the AI describing someone else's version of you.
Entity Consistency Is the Real Lever
Structured data is one part of a larger discipline: making sure every machine-readable description of your organisation says the same thing. Your JSON-LD, your LinkedIn page, your Crunchbase and Wikidata entries where they exist, your G2 and Capterra listings, your press boilerplate, and your own About page should agree on name, founding facts, category, products, and pricing.
Different AI engines weight these sources very differently. Profound's analysis of 680 million citations found Reddit accounted for under 2% of ChatGPT's citations but nearly 7% of Perplexity's, and Semrush's study of 230,000 prompts found the most-cited domains shift sharply from platform to platform and month to month. If your facts differ between your site and a review platform, one model will reproduce your version and another will reproduce the review site's, and your cross-model scores will diverge for reasons that have nothing to do with your product.
Consistency is unglamorous, and it is the single most reliable way to improve how accurately AI describes a brand. Structured data is the part of that job you control completely.
A Note on llms.txt
The llms.txt proposal, a markdown file at your site root summarising your content for AI systems, is often bundled with structured data advice. The evidence here is clearer and less encouraging. Adoption has grown many times over in the past year, to tens of thousands of sites. Yet no major AI provider, including OpenAI, Anthropic, Google, and Perplexity, has confirmed that it reads third-party llms.txt files, and no published study has shown one influencing a citation.
Publishing one costs little. Counting it as a strategy costs more.
A Practical Approach
Use structured data for what it does, and expect nothing more from it.
- Organization schema on every page, with your exact legal and brand names, logo, and sameAs links to your official LinkedIn, Wikipedia or Wikidata entry if you have one, and primary social profiles. This is the anchor for everything else.
- Product or Service schema with names, descriptions, and pricing that match your pricing page word for word. If pricing changes, change it here the same day.
- Article schema on editorial content with author, publication date, and modification date, since freshness is a measurable factor in AI citation and dates are how machines assess it.
- FAQPage and HowTo schema only where a page genuinely answers questions or describes a process. The markup does not create the content's value; it labels it.
- Crawler access that allows the search bots you want to be cited by. OpenAI, Anthropic, and Perplexity each run separate crawlers for training, search, and user-triggered fetches; check that your robots.txt is not blocking the search crawler while trying to block the training one.
Structured data will not get you recommended by AI. Consistent, accurate, machine-readable facts about your organisation reduce the chance that AI recommends a distorted version of you, and structured data is the most controllable place to state those facts. Do it thoroughly, keep it current, and then spend your effort on the content and authority that actually earn citations.
Where This Fits
In a well-run GEO program, structured data sits in the "fix what AI already believes" step, before any new content is created. It is the plumbing, and like plumbing it is noticed mostly when it fails. For the full sequence, see our step-by-step GEO strategy guide.