Schema Markup and AI Citations: What We Tested

The Experiment

We wanted to isolate structured data from everything else that drives citations, so the test was built as matched before/after pairs rather than a loose before-and-after screenshot.

• 50 pages across five industries, paired by topic and roughly matched for authority and word count.
• We added a defined schema stack to each page, left the visible copy unchanged, then removed the markup again on a control run.
• Types tested: Organization (with sameAs), Product/Service, Article/BlogPosting, FAQPage and BreadcrumbList.
• The same governed prompts were re-run on ChatGPT, Perplexity and Gemini, three times each, and coded for mention, citation and absorption depth.

This is a small, directional sample — treat the numbers as signals, not laws. The full rows are in the Schema Markup Impact Data so you can reproduce and extend them.

What Actually Moved

Three effects showed up consistently enough in our 50-page sample to trust as directions.

1. Entity consistency improved first

Organization markup with sameAs links reduced the number of answers that confused a brand with a similarly named company or described it with the wrong category. Identity, not ranking, was the earliest win.

2. Product pages with concrete attributes became more quotable

Product/Service markup that exposed price, availability and key specifications made those facts appear more often in comparative answers — but only when the same values were visible on the page. Structured data made existing evidence easier to lift; it did not invent any.

3. Article and FAQ markup helped extraction, not selection

Once a page was chosen as a source, Article and FAQPage markup correlated with cleaner, more complete absorption of its definitions and steps. We did not see markup alone push an ignored page into the source set — selection was still governed by authority and relevance.

What Didn’t Work

Several common “schema hacks” produced nothing, and some risked penalties.

• Dumping every schema type onto a page with no matching content — irrelevant types were simply ignored.
• Keyword-stuffed or hidden markup that disagreed with the visible page — a mismatch that erodes trust.
• Schema as a substitute for substance — thin pages stayed uncited regardless of how much markup they carried.

This lines up with the absorption research: high-influence pages earn their place through extractable evidence and structure, not metadata alone (arXiv:2604.25707).

The Schema Stack We Recommend

Start small, validate everything, and only mark up what users can actually see.

A minimal starting point

If you deploy nothing else, add a single Organization block site-wide with sameAs links to your verified profiles. This is the cheapest entity-clarity signal in the whole stack.

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "IndexHot",
  "url": "https://indexhot.com",
  "sameAs": [
    "https://x.com/indexhot",
    "https://github.com/indexhot-code"
  ]
}

Validate every block with the Schema Markup Validator before deploying, add it through your CMS template or tag manager rather than hand-editing individual pages, and roll out one type at a time so any change in answers is attributable.

How to Test It Yourself

1. Choose a set of comparable pages and record a baseline with a fixed prompt library.
2. Add one schema type at a time and keep the visible copy unchanged.
3. Wait for recrawl, then re-run the identical prompts on the same platforms and settings.
4. Code mention, citation and absorption separately and compare against the baseline.

The coding rules are in the methodology, and the raw before/after data is free on the Data page.

Frequently Asked Questions