The GEO industry has a schema problem. Not a technical one. A honesty one.
Search any variation of “schema markup for LLM visibility” and you’ll find dozens of articles making confident claims about how structured data gets your content cited by ChatGPT, Perplexity, and Google’s AI Overviews. Some cite specific statistics: 36% more likely to appear in AI summaries, 60% visibility loss without schema, 30% higher citation rates. Almost none of these numbers have a credible source behind them.
The reality is far less certain, far more interesting, and worth understanding properly if you’re making decisions about where to invest your SEO and GEO effort.
TL;DR: There’s one confirmed platform statement (Bing), one null-result empirical study (Search Atlas), zero peer-reviewed research, and a lot of vendor-funded claims. Schema likely contributes to how LLMs build and retrieve knowledge, but the confident “schema gets you cited” narrative is ahead of the evidence. Implement it as infrastructure. Don’t expect it to be a competitive advantage on its own.
Does structured data influence LLM citations? What’s been confirmed
The evidence base for schema’s role in LLM systems is thinner than most people realise.
In March 2025, Microsoft’s Fabrice Canel confirmed at SMX Munich that schema markup helps Bing’s LLMs understand content for Copilot. That’s the only first-party, on-the-record confirmation from a major AI platform. Google’s Search team offered a vaguer nod in April 2025, saying structured data gives an advantage in AI-generated search experiences. OpenAI, Anthropic, and Perplexity have said nothing publicly.
On the research side, a Search Atlas study analysed schema coverage against LLM citation rates across OpenAI, Gemini, and Perplexity. The finding: no correlation. Domains with comprehensive schema performed no better than those with minimal or none. Meanwhile, the 2024 GEO study from Princeton and Georgia Tech, which gets cited constantly in schema discussions, didn’t actually test schema at all. It tested in-content strategies like adding citations and statistics. Those worked. Schema was never part of the experiment.
As of March 2026, there are no peer-reviewed studies on schema’s impact on AI search visibility. None.
The bias problem in GEO and schema research
Before you weigh any of this evidence, consider who produced it.
Schema App, the most vocal advocate for schema’s importance, sells schema implementation as a service. Their revenue depends entirely on schema being valuable. Search Atlas, whose null-result study is the closest thing to empirical evidence, sells an SEO platform that competes with schema-focused tools. Ahrefs, whose recent analysis suggests LLMs read schema as plain text rather than structured data, builds its entire product around backlinks and content metrics. Their business model benefits from a world where links and content matter more than structured data.
Every conclusion in this space conveniently aligns with the commercial interests of whoever published it. That doesn’t make them all wrong, but it should make you sceptical of anyone presenting their findings as settled fact.
How LLMs process schema markup and structured data
To understand where schema might matter, you need to understand that LLM-powered systems aren’t one thing. They’re several different systems with different pathways.
Schema in LLM training data
When an LLM answers from training data alone (ChatGPT without search, Claude without search), it draws on patterns learned during pre-training. The specific handling of schema during training is proprietary and unconfirmed. But the “we don’t know” framing undersells what we can reasonably infer.
LLM training pipelines ingest web crawls. Those crawls contain HTML documents. Those documents have <head> sections containing title tags, meta descriptions, OG tags, and JSON-LD. If you’re building a knowledge store from web documents, you use the metadata that comes with them. Schema.org is a pre-existing, widely adopted, machine-readable framework for declaring what entities exist on a page, how they relate to each other, and what the content is about. It’s already there in the source material. The engineering question isn’t whether to build entity understanding from scratch. It’s whether to use the structured framework that already exists in the data you’re ingesting. It would be a deliberate and somewhat strange decision to strip it out.
We can’t confirm this from outside the labs. But the burden of proof should sit with the claim that training pipelines discard structured metadata, not with the claim that they use it. Discarding it would mean choosing to throw away a signal that every major search engine has found valuable for over a decade.
Schema in RAG and AI search systems
When a retrieval-augmented system answers (Bing Copilot, Perplexity, Google AI Overviews), it fetches web content at query time and passes it to the LLM as reference material. Here, schema’s role is more direct. These systems have search infrastructure that already parses HTML, and the question of whether they retain head content, including JSON-LD, during processing is relevant. Any sensible ingestion system would retain document metadata. The <head> of an HTML document functions like an index card: it tells the system what the document is, who created it, what entities it describes, and how they relate to each other. Discarding that before processing would be like shelving library books without catalogue cards.
Schema and knowledge graphs
Knowledge graphs add another layer. Google’s Knowledge Graph draws from Wikidata, Wikipedia, and structured data across the web. LLM-powered systems that ground themselves in knowledge graphs inherit whatever structured data fed those graphs. This is an indirect but real pathway.
The picture across all pathways
Schema almost certainly contributes in systems with existing search infrastructure (Google, Bing). For LLM training pipelines, it’s unconfirmed but the logic for using it is stronger than the logic for discarding it. And the distinction matters less every month. Every major LLM now has retrieval built in, which means the pathway where schema’s contribution is strongest is becoming the default mode of operation.
One argument that’s hard to dismiss: Google has spent over twenty years trying to algorithmically understand web content without needing publishers to explicitly label it. That’s the entire trajectory from keyword matching to BERT to MUM. They still maintain Schema.org, still encourage adoption, and still build features on top of it. If their algorithms could replace structured data, they would have deprecated it years ago. The fact that they haven’t tells you something about the gap between what algorithms can infer and what explicit structure provides. If Google still needs schema after two decades of development, it seems unlikely that LLM training pipelines, which are younger and solving a harder problem at larger scale, have somehow leapfrogged that limitation.
It’s also worth recognising that schema is just a more explicit version of what semantic HTML already does. An <h2> tells a system “this is a subheading.” A <table> signals tabular data. JSON-LD tells it “this entity is a person, they authored this article, they work for this organisation.” If structured HTML helps LLMs process content (and the research on extraction accuracy suggests it does), it would be odd for its more formalised equivalent to be ignored.
LLM outputs, consensus bias, and what it means for brands
There’s a broader issue worth understanding if you’re thinking about LLM visibility strategically. LLMs learn by statistical pattern matching across their training corpus. If 95% of sources say X and 5% say Y, the model will strongly favour X. That’s how it avoids generating nonsense, but it also means the model inherits the consensus of its training data, including cases where the consensus is wrong, outdated, or superficial.
This creates a feedback loop. LLMs reproduce consensus. People read LLM outputs and create content aligned with that consensus. That content enters future training data, reinforcing the pattern further. Genuinely novel insights that appear in only a few sources get massively outweighed by volume. The outlier who’s actually right gets drowned out.
For brands, this raises a strategic tension. Aligning with consensus increases your probability of being retrieved and cited. Differentiating with original thinking is more valuable but less likely to surface in LLM outputs. LLM visibility currently rewards conformity more than originality.
Does LLM citation actually drive business results?
Most GEO advice assumes a chain: optimise content, get cited by LLMs, gain visibility, drive traffic, generate revenue. But almost nobody is testing the later links in that chain.
There’s very little data on whether being cited by an AI system actually drives meaningful traffic to your site. AI answers are designed to satisfy the query without the user clicking through. That’s the whole point. Studies claiming high conversion rates from LLM referral traffic tend to come from companies that benefit from that narrative, and none of them publish the absolute numbers that would let you assess real business impact.
The more useful framing is probably this: people are spending more time making decisions inside LLM conversations. The path to purchase is changing. The question isn’t whether someone clicks through to your site from an AI citation. It’s whether your brand, your expertise, and your offering are present in the conversation at the point the decision is being shaped. That’s closer to how brand visibility has always worked. You don’t measure a billboard by click-through rate. You measure it by whether you’re in the consideration set when the purchase happens.
Schema markup for LLM visibility: where this actually leaves us
Schema is infrastructure. But it’s not trivial infrastructure. It compensates for a fundamental limitation in algorithmic understanding that hasn’t been solved in over twenty years of trying. It likely contributes to knowledge ingestion and entity resolution in systems with search infrastructure. It probably helps retrieval systems process your content with less ambiguity. And the case for it contributing during LLM training, while unconfirmed, is stronger than the case for it being discarded.
Implement it. The cost is low, the downside is zero, and the potential upside exists across multiple pathways. But don’t let anyone tell you it’s a citation guarantee, a visibility lever, or the difference between being found and being invisible. The evidence doesn’t support those claims. The real competitive advantage comes from what the schema points to: the quality, authority, and distinctiveness of the content itself.
Be sceptical of certainties in this space. The people selling them have products attached.
Get Your Free Copy of JB’s Digital Trends!
This year’s Digital Trends are an absolute cracker. I am seriously pleased with them.
Chock-full of detailed strategic insight, they are an absolutely essential read for any Senior Digital Marketer, C-suite, or Founder who wants to put themselves at the forefront of the game.
Every word, every sentence, is intentional and crafted to within an inch of its meaning.
Know this: I value your time. It won’t be wasted. 🫶
📥 Download your FREE copy today!
Frequently asked questions
Is schema markup a ranking factor for LLMs?
No, and it isn’t a traditional Google ranking factor either. Google has said this explicitly, multiple times. Schema’s value is in helping systems understand your content, not in directly boosting rankings. For LLMs, the evidence is even less clear. One empirical study found no correlation between schema coverage and LLM citation rates.
Do ChatGPT and Claude use schema markup?
Neither OpenAI nor Anthropic has confirmed how their systems handle schema. It may contribute during training data ingestion, and it likely plays a role when these systems use retrieval (ChatGPT Search, for example). But there are no public statements or documented evidence either way.
Which schema types matter most for AI search visibility?
Organisation, Article/BlogPosting, Product, FAQPage, and LocalBusiness are the most commonly recommended. But the honest answer is that nobody has tested which types, if any, influence LLM behaviour. These recommendations are extrapolated from traditional SEO value, not from AI-specific evidence.
Should I implement schema markup for GEO?
Yes. The cost is low, the downside is zero, and it serves multiple purposes across traditional SEO, potential AI visibility, and knowledge graph contribution. But treat it as foundational infrastructure, not as a competitive differentiator. The content it describes matters more than the markup itself.
Is the GEO research on schema reliable?
Very little of it is. There are no peer-reviewed studies on schema and LLM citation. Most published findings come from vendors with commercial interests in the outcome. Apply scepticism to all of it, including studies that conclude schema doesn’t matter.
Does LLM citation actually drive traffic and conversions?
The data is immature. Some studies report high conversion rates from LLM referral traffic, but these tend to come from LLM proponents and lack absolute volume data. The better framing may be that LLM visibility influences the path to purchase by shaping what brands appear in the conversation, rather than driving direct click-through.
