Schema and AI citations: useful, oversold, and not what they sold you
A consultant tells you that adding schema will get you cited by AI. A Google engineer calls schema a stopgap he wishes the systems did not need. A study finds no link between schema and citations at all. They can all be telling the truth — and the gap between them is where most schema advice goes wrong.
the short answer
Schema markup helps AI understand you, but there is no solid proof it gets you cited more. Microsoft says structured data helps its LLMs; Google says it is not a ranking factor and calls it a stopgap; a 2024 study found no schema–citation correlation. What schema reliably does is disambiguate your entity — turning "X leads Y" into a machine-readable fact. Treat it as entity-clarity infrastructure, implement the few core types well, and distrust anyone selling "schema equals more citations."
key takeaways
- The evidence on whether schema moves AI citations is mixed: Microsoft says it helps its LLMs; Google says it is not a ranking factor and calls it a "stopgap"; a 2024 study found no schema–citation correlation.
- What schema does without dispute: it disambiguates your entity — turning "X leads Y" into a machine-readable fact a model can verify.
- LLMs read token streams, not visual layout; clean, structured content extracts more reliably than irregular HTML and messy tables.
- Schema is not a magic citation lever; treat it as entity-clarity infrastructure, not a visibility trick.
- Practical stance: implement the few core types accurately (Organization, Person, Article, FAQPage), and distrust anyone selling "schema equals more citations."
the split verdict on schema and AI — as of 2025
In words, so the table does not carry it alone: as of 2025 the people who would know disagree. Microsoft's Fabrice Canel said at SMX Munich in March that schema helps its LLMs understand content. Google's John Mueller said structured data is not a direct ranking factor, and Gary Illyes called it a stopgap the systems would ideally not need. And an independent December 2024 study found no correlation between schema coverage and citation rates. Four credible voices, four different shades of answer — which is the first clue that the confident "schema gets you cited" pitch is selling more certainty than exists.
Why all four can be right at once
The disagreement is smaller than it looks, because the voices are answering slightly different questions. Canel is saying Microsoft's models can use schema to understand content — a statement about capability, and almost certainly true. Mueller is saying schema is not a direct ranking factor — a statement about Google's ranking systems, also true, and not in conflict with Canel. Illyes is making a wish about an ideal future where machines read language so well that schema becomes unnecessary — a comment on direction, not on whether schema helps today. The study is measuring one specific thing: whether more schema coverage correlates with more citations, and finding it does not. None of these contradicts the others; they only contradict the flattened version a sales deck turns them into.
What gets lost in the flattening is the distinction between comprehension and citation. Schema can genuinely help a model understand your content — what it is, who you are, how your facts relate — without that understanding translating into measurably more citations. Those are different outcomes, and the evidence supports the first far more than the second. A pitch that collapses them, promising citations on the strength of comprehension benefits, is making a leap the data does not. Keep the two separate and the whole topic gets clearer: schema for clarity is well-founded; schema for citations is a hope wearing the costume of a fact.
The schema question, in three parts
What schema definitely does, what the evidence does not show, and why it still earns a place in your stack. Open each layer for the part that changes how you spend.
01 What schema definitely does
Strip away the citation debate and one function is uncontested: schema disambiguates. When you write "John Smith is the CEO of Acme Corp," a human reads the relationship instantly, but a model sees tokens that might relate and might not, with no guaranteed way to verify the claim. Wrap the same fact in Organization schema with a founder property pointing to a Person, and you have turned a probabilistic guess into a machine-readable assertion the system can connect with confidence. That is the durable value — not a ranking signal, a clarity signal. It tells the engine what your page is, who you are, and how your entities relate, in a format built to be read without ambiguity. Everything else about schema is contested; this part is not.
02 What the evidence does NOT show
What the evidence does not support is the headline most schema pitches lead with: add markup, get cited more. A December 2024 study found no correlation between schema coverage and AI citation rates — sites with comprehensive schema did not consistently out-cite sites with little or none. Google has said structured data is not a direct ranking factor and described it as a stopgap. The platforms have not confirmed whether they even preserve schema during crawling or use it at the extraction step. So the specific claim "schema increases your citations" rests on weak ground in 2025, and anyone stating it as settled fact is ahead of the evidence. Holding the claim at its real strength — plausible, unproven — is the honest position.
03 Why it still belongs in your stack
None of that makes schema a waste, which is the overcorrection to avoid. It is cheap, low-risk, and does the disambiguation job reliably, and that job underwrites the entity recognition every engine depends on. Microsoft has confirmed its LLMs use schema to understand content, so for at least one major surface the benefit is stated outright. And clean structured data is part of being legible to systems that read tokens rather than pages. The right posture is unglamorous: implement the core types accurately because clarity is worth having, expect comprehension benefits rather than guaranteed citations, and do not let either the hype or the backlash talk you into treating a cheap, sensible practice as either a silver bullet or a scam.
The deeper reason structure matters: models read tokens, not pages
To see why structured data helps comprehension even where it does not move citations, it helps to remember what a model actually consumes. An LLM does not see your page the way a person does — the layout, the visual hierarchy, the design cues that tell a human what is a heading and what is an aside. It sees a stream of tokens and the relationships it can infer between them. Research on document understanding has shown that when models ingest raw, irregular HTML or tangled tables, reliability drops; order, formatting and structure materially affect how well the model can extract what a page means. Structure is not decoration to a model; it is the difference between a page it can parse cleanly and one it has to guess at.
Schema is one way to hand the model that structure explicitly, but it is not the only one, and this is where a lot of schema advice quietly overreaches. A page written in clean, declarative prose, with clear headings, short self-contained passages and facts stated plainly, is already far more extractable than a wall of hedged marketing copy — with or without markup. Schema adds a machine-readable layer on top of that; it does not rescue a page that is mush underneath. The sequence matters: get the underlying content clear and structured for extraction first, then add schema to label it. Markup on top of muddle helps far less than the order reversed.
Which types are worth your time
The temptation, once you accept schema is worth doing, is to mark up everything — and that is the wrong instinct. A short list carries almost all the value. Organization and Person, linked together, establish who you are and who speaks for you; this is the entity backbone, and getting it consistent across your site and the wider web is the single highest-value piece. Article or BlogPosting marks your content and attributes it to an author. FAQPage turns question-and-answer content into discrete pairs a model can lift cleanly. WebPage states what a given page is for. Those few, done accurately, cover the cases that matter for AI comprehension.
Beyond that short list, returns fall off fast and risks rise. Every additional type is more to maintain and more that can drift out of sync with what is actually on the page — and schema that contradicts the visible content is worse than no schema, because it teaches the model something false about you. The discipline is to mark up what you genuinely are, accurately, and stop. A handful of correct, consistent types beats an exhaustive library of markup that nobody keeps current. Accuracy and consistency are the goals; completeness is a vanity metric that costs maintenance and buys little.
How to do it without overpaying
Because schema attaches so easily to a citation promise, it has become a line item some consultancies price well above its worth. Be clear-eyed about the economics. Implementing the core types correctly is well-documented work: a competent developer can do it, generators can scaffold it, and validation tools can check it. It does not, for most sites, require an expensive specialist engagement. The fees tend to climb precisely when the pitch ties schema to a guarantee of more AI citations — and that guarantee is exactly the part the evidence does not support. If a proposal is priced against citation lift, you are being charged for an outcome no one can promise.
There are real exceptions: genuinely complex entity modelling, large catalogues, knowledge-graph work where expert help earns its cost. But the median site is not that. For most, the right spend is modest — implement the core types accurately, keep them consistent, validate them, and move on — and the money saved is better aimed at the things that actually move AI visibility: being genuinely worth citing, present where your category is discussed, and clear enough to extract. Schema is table stakes done cheaply, not a budget line that earns its keep by the size of the invoice.
The honest verdict
Put the pieces together and the verdict is unglamorous, which is usually a sign it is right. Schema is worth doing, for clarity rather than for citations. It reliably disambiguates your entity and helps models that read tokens make sense of your content, and at least one major engine confirms its LLMs use it. It is not a proven citation lever, and the confident pitches that say otherwise are running ahead of the evidence as it stood in 2025. The trap is the binary: hype on one side selling schema as a visibility cheat code, backlash on the other dismissing it as snake oil. Neither is right. It is a cheap, sensible piece of infrastructure that earns its small place by making you legible — no more, and no less.
The AC Group has spent 27 years separating what a technique actually does from what it is sold as doing, and schema is a clean example of the gap. Implement it well because clarity compounds and the cost is low; do not buy it as a citation guarantee, because that guarantee does not exist. When the proof arrives one way or the other — and the studies are coming — we will follow it. Until then, the disciplined move is to do the cheap, clear thing and spend the saved budget where the evidence already points.
Schema and AI: quick answers
If schema does not guarantee citations, why bother with it?
Because citation lift is not the only thing it buys, and the other things are real. Schema reliably does one job no one disputes: it disambiguates your entity, turning "John Smith leads Acme" from a string a model has to guess about into a machine-readable fact it can connect with confidence. That clarity helps every engine recognise who you are, which is the foundation the rest of AI visibility sits on, even where it does not move a citation count directly. It is also cheap and low-risk to implement well. The honest framing is not "schema gets you cited" — the evidence for that is weak — but "schema makes you legible," and legibility is worth having on its own terms. Bother with it for clarity; do not bother with it because someone promised you citations.
Which schema types matter most for AI?
A short list does most of the work, and adding more rarely helps. Organization and Person schema, linked together, establish who you are and who speaks for you — the entity backbone. Article or BlogPosting marks up your content and its author. FAQPage structures question-and-answer content into discrete, extractable pairs. WebPage defines the purpose of a page so a model can categorise it. Those few, implemented accurately and kept consistent with what is actually on the page, cover the cases that matter. The mistake is marking up everything indiscriminately, which adds maintenance burden and risk of mismatch without adding clarity. Pick the two or three types that describe what you genuinely are, get them exactly right, and stop. Comprehensive is not the goal; accurate and consistent is.
Does Google use schema for AI Overviews?
Not as a confirmed citation lever, and Google has been notably cool on the idea. John Mueller has said structured data is not a direct ranking factor, and Gary Illyes has described schema as a stopgap — useful while machines cannot read language cleanly, but something an ideal system would not need. Microsoft has been the opposite, confirming that schema helps its LLMs understand content for Copilot. So the answer splits by engine and stays genuinely uncertain: the platforms have not publicly confirmed whether they preserve schema during crawling or use it for extraction, and at least one 2024 study found no correlation between schema coverage and citation rates. Treat schema as helping comprehension broadly rather than as a Google-AI-Overviews trick, because the evidence for the latter is not there.
Is schema worth paying a consultant a large fee for?
Rarely, and the pitch should make you cautious. Basic, correct schema for the few types that matter is a well-documented task that a competent developer can implement and a generator can scaffold; it does not require an expensive specialist engagement to do well. The fees tend to climb when the pitch attaches schema to a promise of more AI citations — a promise the evidence does not support. If a proposal is priced against citation lift, you are paying for an outcome no one can guarantee. There are genuinely complex entity and knowledge-graph projects where expert help earns its cost, but for most sites the right spend is small: implement the core types accurately, keep them consistent, and put the money you saved into the things that actually move AI visibility, like being worth citing in the first place.
A note on sources and certainty
The voices here are real and on the record: Microsoft's Fabrice Canel at SMX Munich in March 2025, Google's John Mueller and Gary Illyes across 2025, and an independent December 2024 study finding no schema–citation correlation. We present the topic as genuinely unsettled because it is — the platforms have not fully disclosed how they use schema, and the studies so far point more at comprehension than at citation lift. We have kept to what was knowable as of this writing; sharper studies are in progress and may move the verdict in either direction, and we will say so plainly when they land. What is not in dispute is the narrow, durable claim: schema disambiguates entities and aids extraction, which is reason enough to do the cheap version well, and reason enough to refuse the expensive version sold on a promise the evidence has not yet earned.