Does schema markup help AI search?
We sell this work and we publish JSON-LD on every page here. The evidence that it drives AI citations is weak, and pretending otherwise would be the easier answer.
Does schema markup help AI search?
There is no good evidence that adding schema.org markup increases how often AI assistants cite a page. Google states in its documentation that structured data is not required for generative AI search and that no special schema.org markup needs to be added. A matched-control study of 1,885 pages that newly added JSON-LD found changes indistinguishable from noise. A controlled platform test found six of seven AI systems could not fetch or correctly interpret schema when asked to. The defensible reason to publish structured data is that it is the precise way to state facts about a business, and that it still earns rich results in classic search. That is a smaller claim than the one usually sold.
Our position, stated first, because we have an interest
This company sells AI search optimization. Every page on this site carries JSON-LD. If schema markup were a proven citation lever we would have every commercial reason to say so.
It is not proven, and the strongest available evidence points the other way. We are going to lay that out before we explain why we still ship it, because a page that defends its own deliverable past the evidence is worth nothing as a reference.
What Google documents
Google addresses this directly in two separate places, and the answer is deflationary both times.
The generative AI optimization guide lists "Overfocusing on structured data" as a mistake and says: "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add."
Google supplies the balancing clause in the same breath, and it matters: "However, it's a good idea to continue using it as part of your overall SEO strategy, as it helps with being eligible for rich results on Google Search."
So Google's documented position is that structured data buys eligibility for rich results in classic search, and is not a lever for AI Overviews or AI Mode specifically. That is a narrower and more useful statement than either camp usually quotes.
The three tests that matter
Ahrefs tracked 1,885 pages that newly added JSON-LD between August 2025 and March 2026, against roughly 4,000 control pages that never added any, measured across AI Overviews, AI Mode and ChatGPT. AI Mode moved 2.4 percent, ChatGPT 2.2 percent, both indistinguishable from noise. AI Overviews moved 4.6 percent in the negative direction, statistically significant, and Ahrefs explicitly declined to attribute that to the markup.
Ahrefs also named the confound that produces the folk belief. Cited pages are roughly three times more likely to carry JSON-LD, because schema lives on well-maintained sites that are doing everything else well too. That is a selection effect, not a mechanism.
Otterly.ai ran a controlled experiment across seven AI platforms and 319 monitored prompts between December 2025 and March 2026, with competitor coverage tracked as a drift control. Six of the seven platforms could not fetch or correctly interpret schema markup when asked to directly. Only Gemini could. No measurable citation effect appeared.
Mark Williams-Cook ran what he called the Duck Test: deliberately invalid JSON-LD, with a made-up @type and made-up properties. ChatGPT and Perplexity extracted the facts anyway, and Perplexity claimed it had used structured data. If a model reads invalid JSON-LD as happily as valid JSON-LD, it is not parsing it as data. It is reading it as slightly oddly punctuated prose.
That last result is the one that reframes the subject. The question stops being "does the model parse your schema" and becomes "does the fetch tool strip script tags or not."
The strongest objections to that evidence, stated fairly
The Ahrefs study has a real sampling limit that Ahrefs disclosed: every page in it already had more than 100 AI Overview citations before the markup was added. It therefore tests whether schema lifts pages already inside the consideration set. It does not test whether schema helps an unknown entity get disambiguated in the first place, which is the case local vendors actually sell against.
The Otterly experiment ran on a single SaaS brand and says so. It is not a local business test, and the LocalBusiness markup question has never been tested on actual local businesses by anyone.
John Mueller, asked directly whether schema helps language models, answered "yes, no, and it depends," which is not a denial. The only affirmative platform-side statement anyone has is a conference remark attributed to Microsoft, never published in Microsoft documentation, and contested as to what it was actually about.
So the honest summary is not "schema does nothing." It is that the one matched-control test found nothing, the one platform fetch test found six of seven systems could not read it, nobody has tested the disambiguation case, and no platform has ever documented using it for AI answers.
Why we still publish it
Three reasons, none of which is a citation claim.
First, it is the accurate way to state facts about an entity. Writing the business name, the former name, the founder, the address and the hours as structured properties forces those facts to be consistent across a site, and inconsistency is a real problem we can observe directly. Gaetano Pizzi put the defensible version well: markup does not create visibility, it helps ensure a business is represented accurately when it is surfaced. That claim is reasonable and it is also uncited, and we label it as reasoning rather than as evidence.
Second, it still does the classic-search job Google describes. Rich result eligibility is documented, and it is the benefit Google itself points at.
Third, the cost is close to zero once it is generated from the same data the pages render from. A cheap thing with a documented secondary benefit and an unproven primary one is worth doing, as long as nobody is invoiced for the unproven part.
What we will not do is tell a customer that adding markup will get them cited. Nobody has shown that, and the direction of travel in Google's own documentation is toward fewer supported structured data types, not more. Seven types were retired in June 2025, and FAQ rich results stopped appearing entirely on 7 May 2026.
| Claim | Status | Basis |
|---|---|---|
| The markup is valid schema.org | True and easy to verify | The vocabulary is public and a validator will confirm it |
| It produces a Google rich result | True for some types, shrinking | Google documents an explicit and contracting list of supported features |
| It causes an AI system to cite you | Not established by anyone | One matched-control study found noise; no platform documents using it |
What is actually established, and how
Sorted by how strong the evidence is, not by how convenient it is.
| Claim | Basis |
|---|---|
| Google states structured data is not required for generative AI search and that no special schema.org markup needs to be added. | Documented by the platform |
| Google states structured data remains worth using because it supports eligibility for rich results in classic Search. | Documented by the platform |
| A matched-control study of 1,885 pages adding JSON-LD against about 4,000 controls found effects indistinguishable from noise on AI Mode and ChatGPT. | We measured this |
| Six of seven AI platforms could not fetch or correctly interpret schema markup when asked directly, in a controlled 319-prompt test. | We measured this |
| Models extracted facts from deliberately invalid JSON-LD, which suggests the block is being read as text rather than parsed as data. | We measured this |
| Google has retired supported structured data types repeatedly, including seven in June 2025 and FAQ rich results on 7 May 2026. | Documented by the platform |
| Whether markup helps an unfamiliar entity get disambiguated, which no published study has tested. | Not publicly documented |
What nobody can currently tell you
Stated because the alternative is implying a certainty that does not exist.
Nobody has run a schema test on actual local businesses. The only controlled platform experiment used a single business-to-business software brand, and its authors say so.
The disambiguation case is untested. Every published study measured pages that were already being cited, so none of them speaks to a business the systems do not yet know.
Whether any AI system uses structured data internally is undocumented in both directions. No platform has stated that it does, and no platform has stated that it does not.
Whether the negative movement Ahrefs measured on AI Overviews means anything at all. Ahrefs declined to attribute it and we decline as well.
Whether the fetch layer matters more than the markup. The invalid-JSON-LD result suggests the deciding variable is whether a fetcher strips script tags, and that has not been measured across platforms at scale.
What people get wrong about this
-
Adding JSON-LD improves your odds of being cited by AI.
What is actually the case
The only matched-control test of that exact question returned a null result, and Google documents that no special markup is needed. The observed association between markup and citations is explained by markup appearing on better-maintained sites.
-
Cited pages have schema three times more often, so schema is working.
What is actually the case
That ratio comes from the same study that found no effect when schema was added. It is the textbook shape of a confounded correlation, and the researchers who reported the ratio said so.
-
More markup types means more AI visibility.
What is actually the case
Google has been removing supported types for years rather than adding them, and no additional type has ever been documented as an input to an AI answer.
-
FAQPage markup earns you an FAQ rich result.
What is actually the case
FAQ rich results stopped appearing in Google Search on 7 May 2026 and the documentation was removed. The markup remains valid vocabulary that produces no search feature.
How to check this on your own site
You should not have to take our word for any of it.
- Fetch one of your own pages with a plain text client and confirm the JSON-LD survives. Then check whether the facts a reader would need appear in the visible HTML as well, because the extraction evidence suggests visible text is what actually gets read.
- Ask an assistant with browsing enabled a factual question about your business whose answer exists only inside your JSON-LD and nowhere in the visible copy. Record whether it gets it right. That is the experiment Otterly ran and you can run a small version of it in ten minutes.
- Validate the markup, then separately check the current rich results gallery to see whether the type you used still produces any documented search feature.
Questions we get asked constantly
Should I remove the structured data I already have?
No. Google's standing position is that unused structured data causes no problems for Search. There is no benefit to stripping it, and it still carries the accuracy and rich-result rationale.
Does LocalBusiness markup help a local business get recommended?
Unknown. It is the single most commonly sold local schema deliverable and it has never been tested on local businesses by anyone who published a method.
Why do so many guides say the opposite?
The disconfirming evidence is recent, arriving mostly between late 2025 and early 2026, and much of it was published by vendors against their own sector interest. The pages ranking for this question were written before it existed and have not been revised.
Where this came from
Every factual claim above traces to one of these. Each entry says what it supports and the date it was read, because platform documentation changes without notice.
-
Google AI optimization guide, structured data section — Google Search Central
The overfocusing-on-structured-data guidance, including both the no-special-markup sentence and the rich-results clause that follows it.
Primary source · read 2026-09-03
developers.google.com/search/docs/fundamentals/ai-optimization-guide
-
AI features and your website — Google Search Central
Second Google statement that no special schema.org structured data needs to be added for AI features.
Primary source · read 2026-09-03
-
Schema markup and AI citations: a matched-control study — Ahrefs
1,885 pages newly adding JSON-LD versus roughly 4,000 controls; AI Mode 2.4 percent, ChatGPT 2.2 percent, AI Overviews minus 4.6 percent; and the confound that cited pages are about 3x more likely to have JSON-LD.
Primary source · read 2026-09-03
-
The real impact of schema markup on AI search — Otterly.ai
Seven platforms, 319 monitored prompts, December 2025 to March 2026: six of seven could not fetch or correctly interpret schema, and no measurable citation effect was found.
Primary source · read 2026-09-03
-
Schema, LLMs, and the low bar for evidence — Mark Williams-Cook
The Duck Test: models extracted data from deliberately invalid JSON-LD with a fabricated type, indicating text extraction rather than structured parsing.
Primary source · read 2026-09-03
markwilliamscook.substack.com/p/schema-llms-and-the-low-bar-for-evidence
-
The Ahrefs schema study is right, and it is testing the wrong thing — I Love SEO
The sampling critique: every studied page already had 100+ AI Overview citations, so the disambiguation case was not tested.
Secondary source · read 2026-09-03
iloveseo.net/the-ahrefs-schema-study-is-right-and-its-testing-the-wrong-thing/
-
Schema and local visibility in Google AI — Search Engine Land, Gaetano Pizzi
The defensible pro-schema position: markup does not create visibility, it supports accurate representation when a business is surfaced.
Secondary source · read 2026-09-03
searchengineland.com/schema-local-visibility-google-ai-470906
-
Simplifying the search results page — Google Search Central Blog
Seven structured data types phased out in June 2025, evidencing a contracting rather than expanding supported set.
Primary source · read 2026-09-03
developers.google.com/search/blog/2025/06/simplifying-search-results
-
Search Central documentation updates changelog — Google Search Central
FAQ rich results stopped appearing in Google Search on 7 May 2026 and the documentation was subsequently removed.
Primary source · read 2026-09-03
-
Mueller on whether schema helps LLMs — Search Engine Roundtable
John Mueller answering "yes, no, and it depends" when asked directly whether schema helps language models.
Secondary source · read 2026-09-03
www.seroundtable.com/mueller-schema-helps-llms-google-40693.html