Schema and AI citations: what the evidence supports

CH.01

Start with the system and the measured outcome

Website markup, enterprise knowledge graphs and AI citations are different subjects. A result from one system cannot establish the effect of adding JSON-LD to a public website.

Two original studies illustrate the distinction. data.world tested question answering over an insurance database and a knowledge graph. Ahrefs tracked citation changes on pages that introduced JSON-LD. Their treatments, samples and outcomes differ, so their numbers should not be combined into a schema return forecast.

CH.02

The 16.7% and 54.2% result concerns database queries

data.world’s November 13, 2023 benchmark asked GPT-4 to produce SQL queries or SPARQL queries over a knowledge graph. It evaluated 43 insurance-domain questions. This was not a test of adding schema.org markup to a webpage.

The knowledge-graph setup supplied an ontology and mappings for the enterprise data. That context is materially different from publishing an Article or Organization object in a webpage. The study does not establish that JSON-LD prevents an assistant from misstating a price or author.

If you operate an internal question-answering system, test it against your own reference answers and data model. For a public website, use the structured-data requirements of the intended consumer and check the rendered facts independently.

CH.03

What the Ahrefs citation study can tell us

Ahrefs observed 1,885 pages introducing JSON-LD between August 2025 and March 2026 and compared them with 4,000 controls. Its matched analysis found a 4.6% relative AIO decline, while the AI Mode and ChatGPT estimates were not statistically distinguishable from zero.

The pages already had substantial AIO citation activity. The study pooled schema types, used a 30-day post-treatment window and could not isolate other changes made at the same time. Its findings do not settle whether a specific markup type helps an uncited page or a different consumer.

CH.04

Describe the actual organization and author

Organization and Person markup can describe who a page is about and who wrote it. Google documents Organization markup as a way to provide administrative details and help disambiguate an organization. This is not a guaranteed route into a knowledge graph or an AI answer.

Use the real name, official URL and verifiable relationships. A sameAs link identifies another page about the same entity; it must not imply a partnership or identity that does not exist. Markup should agree with visible page content.

CH.05

Allocate implementation work by its documented purpose

Begin with the page’s content and the system that will consume the markup. Prioritize a supported search feature, an identity clarification or a known data mismatch that you can verify.

Google says no special schema.org markup is required for AI Overviews or AI Mode. Pages still need ordinary Search eligibility and snippet eligibility. Keep useful structured data accurate, but do not treat extra types as a shortcut to AI visibility.

CH.06

Review the change with the appropriate evidence

Keep implementation validity, factual accuracy and search outcomes separate. This prevents a successful validator result from being reported as a business or citation gain.

Choose a few representative pages and record what changed. Retain the public facts, generated markup, validation result and observation date so another reviewer can reproduce the check.