Why the schema claim sounds convincing.
Schema helps machines understand information. AI search relies on machines understanding information. The conclusion feels obvious, but the causal link is not.
Ahrefs initially analysed six million URLs and found that pages cited by AI were almost three times more likely to contain JSON-LD than pages that were not cited.[1]
- 01
AI-cited pages use schema more often
ObservedThe association in the data is real and large.
- 02
Schema appears on technically mature sites
ConfoundersWell-maintained sites implement it as a matter of course.
- 03
Those sites also have stronger content and links
Alternative causesAuthority, freshness and depth travel together with technical maturity.
- 04
Therefore adding schema will increase citations
Not establishedThis step does not follow from the three above it.
The question for an SEO leader is not whether cited pages often contain schema. It is whether adding schema to an existing page causes its citations to increase.
What structured data actually does.
Structured data provides explicit, machine-readable information about the content and entities on a page. Google describes it as a standardised format for classifying page content and giving Search explicit clues about meaning, and generally recommends JSON-LD because it is easier to implement and maintain.[2]
Its established purpose in Google Search is rich-result eligibility. Eligibility means the page can be considered for a supported presentation, not that the feature will appear.
| Layer | What schema does | What it does not do |
|---|---|---|
| Description | Labels entities and properties | Does not make the information useful |
| Understanding | Gives systems additional clues | Does not guarantee ranking |
| Eligibility | Supports defined search features | Does not guarantee appearance |
| Consistency | Expresses facts in a standard format | Does not prevent another source being selected |
| AI visibility | May support interpretation | Does not guarantee citation |
Machine readability and source selection are different outcomes.
What 1,885 pages adding JSON-LD showed.
To move beyond correlation, Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026. The pages were matched with roughly 4,000 controls, and citation changes were measured for 30 days either side of implementation across three AI surfaces.[1]
The AI Overview decline should not be read as proof that schema damages citations. Both treatment and control groups were already declining, and co-occurring changes may have affected treated pages.
What the experiment proves, and does not.
A matched experiment is more useful than a simple comparison, but it still has boundaries.
What the evidence shows
- Adding JSON-LD did not generate a reliable citation increase in this study
- Correlation between schema and citations is not causal uplift
- Platform-wide changes need controls to interpret
What it does not show
- That schema has no value
- That schema reduces AI visibility by 4.6%
- That every schema type behaves the same
- That the result holds forever, or on every platform
Schema types were pooled
Article, FAQ, Product, HowTo and Organisation may behave differently.
A thirty-day window
Longer crawling or processing cycles could produce different results.
Other changes coexist
Content, links, templates and technical fixes often launch together.
Already-cited pages
The dataset is strongest for uplift among pages already in the consideration set.
The study tested JSON-LD during a defined period. It did not test every format, implementation or future platform behaviour.
Google does not require schema for AI.
Google’s generative AI guidance says structured data is not required for generative search, that there is no special schema.org markup for AI Overviews or AI Mode, and that existing structured data should continue to serve established Search features.[3]
No additional AI markup
A page must be indexed and eligible to appear in Search with a snippet. Special schema is not required.[4]
Grounding, not a guarantee
Structured data may support clearer grounding, but it does not guarantee visibility or grounding traffic.[5]
Structured data can describe information. Retrieval and citation systems still decide whether the information is relevant enough to use.
Should you keep implementing schema?
Yes, where schema has a defined and supported purpose: rich-result eligibility, product or software information, organisation details, article metadata, breadcrumbs, reviews, events, jobs and consistent machine-readable facts.
Keep the visible page as the source of truth, and implement in this order.
- 01
Visible information
What the reader can actually see and verify on the page.
- 02
Supporting evidence
The data, examples and expertise behind the claims.
- 03
Structured description
Markup that labels what is already there.
- 04
Feature eligibility
The supported presentations the markup unlocks.
A Product object cannot compensate for missing product information. Article markup cannot make generic writing original. Organisation markup cannot create authority the business has not established elsewhere.
Measure schema across four outcomes.
Do not collapse these into one number. Each answers a different question, and only the last one is about AI.
| Outcome | The question it answers | What to track |
|---|---|---|
| Validity | Is the markup technically correct? | Valid items · errors · warnings |
| Eligibility | Can the page qualify for a feature? | Eligible pages · supported types |
| Search presentation | Does the enhanced result appear? | Rich-result impressions · clicks · CTR |
| AI citation | Did citations change after implementation? | Citation rate · cited pages · citation share |
Test citation effects as a hypothesis.
If you want to know whether schema moves citations on your own site, run it as an experiment rather than as an assumption.
- 01
Define treatment and control groups
Use comparable pages with stable pre-test citation levels.
- 02
Record deployment and co-occurring changes
Document schema, content, template and technical release dates.
- 03
Hold prompts and surfaces constant
Track the same query set, and report AI Overviews, AI Mode, ChatGPT and Copilot separately.
- 04
Measure long enough to reduce noise
Compare whether treated pages outperform controls, not whether both groups moved during a platform-wide shift.
Schema describes. It does not select.
Schema remains useful technical infrastructure. It can classify information, clarify entities and make pages eligible for supported search features. Those are meaningful outcomes.
Current evidence does not show that adding JSON-LD reliably increases citations in Google AI Overviews, Google AI Mode or ChatGPT.
Schema helps describe the source. It does not guarantee that the source will be selected.
A system may understand exactly what a page contains and still decide that another source is more relevant, current, authoritative or useful for the answer.
Citations and primary sources.
- [1]We tracked 1,885 pages adding schemaAhrefs · Matched analysis across AI Overviews, AI Mode and ChatGPTOpen source ↗
- [2]Introduction to structured data markupGoogle Search Central · JSON-LD, classification and rich-result eligibilityOpen source ↗
- [3]Optimising for generative AI featuresGoogle Search Central · No special schema required for generative searchOpen source ↗
- [4]AI features and your websiteGoogle Search Central · Technical eligibility for AI search featuresOpen source ↗
- [5]Bing Webmaster GuidelinesMicrosoft Bing · Structured data, grounding and visibility limitationsOpen source ↗