What schema markup actually does for AI.
Schema markup is sold hard as the thing that gets you cited by AI. I implement it for every client. I also think most of what is claimed about it is unproven, and I would rather you heard that from me than worked it out later.
Here is what is actually established, what is inference, and what is marketing — and why the honest version still ends with you implementing schema.
What is actually confirmed
One first-party confirmation exists, and it is worth having precisely. At SMX Munich in 2025, Microsoft's Fabrice Canel said that schema markup helps Bing's language models understand content for Copilot. Google has acknowledged in general terms that structured data is advantageous in AI search without being specific about mechanism.
OpenAI, Anthropic and Perplexity have said nothing publicly about using schema.
That is the complete list of what the companies themselves have stated. One clear confirmation, one vague one, three silences.
What the research shows
Thinner than the industry implies. As of March 2026 there are no peer-reviewed studies on schema's impact on AI search visibility.
A Search Atlas study found no correlation between schema coverage and citation rates across OpenAI, Gemini and Perplexity. The 2024 Princeton and Georgia Tech study that gets cited constantly in GEO marketing did not test schema at all — it tested content strategies like adding citations and statistics. It is cited as schema evidence by people who did not read it.
The distinction that actually matters
Almost every argument about this collapses because two different claims get treated as one.
The direct claim: a language model reads your JSON-LD and is more likely to cite you because of it. Unconfirmed. No provider has described this mechanism and no study demonstrates it.
The indirect claim: schema helps search infrastructure parse and classify your content, and AI assistants retrieve from that infrastructure. Well supported. Bing has confirmed it. Retrieval-augmented systems query search indexes, and those indexes demonstrably use structured data.
The indirect claim is the one that justifies the work. It is less exciting, which is why the direct one gets sold instead.
| Claim | Status | What supports it |
|---|---|---|
| Helps Bing and Copilot understand content | Confirmed | Microsoft, on the record |
| Helps search indexes classify you | Established | Rich results, knowledge panels, entity resolution |
| Helps AI assistants that retrieve from search | Strong inference | Follows from the two above |
| Resolves who you are across sources | Established | Entity disambiguation is what @id is for |
| LLMs parse your JSON-LD directly | Unconfirmed | No provider statement, no study |
| Schema coverage raises citation rates | Contradicted | Search Atlas found no correlation |
Why I still implement it on every project
Because the proven benefits are sufficient on their own, and because one of them is the entire job.
Entity resolution. This is the one that matters most and gets discussed least. Schema lets you state, in a form a machine cannot misread, that the business on your site is the same business as the one on your Business Profile, your LinkedIn and every directory listing. You do that with a stable @id referenced from every page, and sameAs pointing at your external profiles.
When those facts agree, a machine can assert things about you confidently. When they disagree, it hedges or omits you — and omission is what being invisible actually looks like.
Rich results. Established, visible, and still worth having in ordinary search.
It costs almost nothing once. Static JSON-LD in the head. No runtime cost, no maintenance beyond keeping it true.
So: implement it. Just implement it for entity resolution and search infrastructure, which are real, rather than for direct LLM parsing, which nobody has demonstrated.
What to implement, in order
1. One Organization or LocalBusiness node with a stable @id. Something like https://yoursite.com/#business. Define it once. Reference it by @id from every other page rather than redefining it. Two definitions of the same company is the contradiction you are trying to avoid.
2. One Person node, same discipline, if a named individual is part of the brand. worksFor points at the business @id.
3. sameAs on both, listing every external profile you control. This is the line that does entity resolution. Without it a machine has no way to know your LinkedIn is you.
4. Page-level types that are true. Article or BlogPosting on articles, with author and publisher referencing the @ids above. Service on service pages. Nothing aspirational.
5. FAQPage only where the answers are visible on the page. Google requires the markup to correspond to content a visitor can read. Marking up questions that exist only in the JSON is a guidelines violation, and it is common — I have found it on sites that were otherwise carefully built, including ones I inherited.
Two schema anti-patterns worth naming. Do not self-apply aggregateRating or review to your own business; Google treats self-serving review markup as ineligible. And do not let your schema drift from your visible copy — a page claiming one address in markup and another in text is worse than a page with no markup.
How to tell if yours is working
Not by counting schema types. By checking whether a machine can answer questions about you correctly.
Ask an assistant who runs your business, where it is located, and what it does. If the answer is right, your entity is resolving. If it hedges, confuses you with someone similar, or declines, you have a contradiction somewhere, and the schema is where to start looking because it is the part you fully control.
That test takes two minutes and tells you more than any schema validator, because it measures the outcome rather than the implementation.
Run that test properly. The free audit asks a live assistant real questions about your business and shows which sources it used to answer.
Run the free auditCommon questions
Does schema markup help LLMs cite my content?
Indirectly, yes; directly, unproven. Microsoft has confirmed that schema helps Bing's language models understand content for Copilot, and search indexes demonstrably use structured data, so assistants that retrieve from those indexes benefit. But no provider has described a language model parsing your JSON-LD directly, and a Search Atlas study found no correlation between schema coverage and citation rates across OpenAI, Gemini and Perplexity.
Is there research proving schema improves AI visibility?
No. As of March 2026 there are no peer-reviewed studies on schema's impact on AI search visibility. The 2024 Princeton and Georgia Tech study frequently cited as evidence tested content strategies such as adding citations and statistics, not schema markup.
If the direct benefit is unproven, why implement schema at all?
Because the established benefits justify it on their own. Schema drives rich results in ordinary search, it is confirmed to help Bing and Copilot, and most importantly it performs entity resolution — letting a machine establish that the business on your site is the same entity as the one on your Business Profile and directory listings. Consistent facts are what allow a model to assert something about you rather than hedge.
What schema types matter most for a business?
A single Organization or LocalBusiness node with a stable @id referenced from every page; a Person node for any named individual in the brand, with worksFor pointing at the business; sameAs on both listing every external profile you control; truthful page-level types such as Article and Service; and FAQPage only where the answers appear as visible copy on the page.
Can schema markup hurt my site?
Yes, in three ways. Self-applied aggregateRating or review markup on your own business is treated by Google as self-serving and ineligible. FAQPage markup whose questions do not appear as visible page content violates Google's structured data guidelines. And schema that contradicts your visible copy or your other listings creates exactly the inconsistency that causes a model to decline to assert anything about you.
Sources
- Belmore Digital — Does schema markup help LLMs? What the evidence actually shows (provider statements, Search Atlas null result, absence of peer-reviewed studies)
- Schema.org — the vocabulary itself
- Google — structured data features and eligibility requirements
About the author. Siamak Kalhor is an AI marketing and LLMO consultant in Los Angeles with more than forty years of experience in marketing strategy. He advises professionals and e-commerce leaders on becoming the answer AI assistants give. Get in touch, or connect on LinkedIn.