Schema Markup & AI Citations: What Studies Show

For years, schema markup (structured data, usually in JSON-LD format using schema.org vocabulary) has been treated as a near-essential SEO tactic. Marketers added it hoping for ranking boosts, richer search results, and, more recently, higher visibility in AI answers from Google AI Overviews, ChatGPT, Perplexity, Gemini, and similar systems. The common claim in generative engine optimization (GEO) and answer engine optimization (AEO) circles has been that schema helps AI systems “understand” pages better and cite them more often.

Recent controlled studies and platform statements challenge that assumption, especially for AI citations. Correlation exists, but causation does not appear to hold in the data we have.

Correlation Is Real, Causation Is Not

An Ahrefs analysis of roughly 6 million URLs found that pages cited by AI were nearly three times more likely to include JSON-LD schema than non-cited pages. About 53% of AI-cited pages carried some form of schema.

This statistic circulated widely as evidence that schema drives AI visibility. However, well-maintained, authoritative sites tend to implement technical best practices (including schema) and produce high-quality content, earn the links, and build the brand signals that AI systems favor. Schema often rides along with those stronger signals rather than causing the citations.

The Key Controlled Experiment: Ahrefs (1,885 Pages)

To test causality, Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026. Each was matched against control pages from different domains that had similar pre-treatment citation levels but never added schema. Citations were measured 30 days before and after the change across Google AI Overviews, Google AI Mode, and ChatGPT using difference-in-differences and other statistical tests.

Results:

• Google AI Overviews: -4.6% (small relative decline)
• Google AI Mode: +2.4% (statistically indistinguishable from zero)
• ChatGPT: +2.2% (statistically indistinguishable from zero)

Multiple tests pointed to the same conclusion: adding schema produced no meaningful citation uplift. The pages studied were already heavily cited (100+ AI Overview citations beforehand), so the finding applies specifically to pages already visible to AI systems. It does not fully answer whether schema helps obscure pages get noticed in the first place.

Supporting experiments (including one cited by Ahrefs) show that major AI systems, when fetching pages in real time, primarily extract visible HTML content and largely ignore JSON-LD script tags.

Additional Independent Analyses

Search Atlas examined schema coverage at the domain level against visibility scores in OpenAI, Gemini, and Perplexity responses. Domains with full schema coverage performed no better than those with minimal or no schema. Visibility distributions looked nearly identical across schema adoption levels.

SE Ranking analyzed 129,000 domains for ChatGPT citation factors. Strongest predictors were referring domains, domain trust, traffic, content depth, freshness, expert quotes, and statistical data. FAQ schema actually underperformed: pages with it averaged fewer citations than those without.

Semrush focused on visible text qualities associated with AI citations. Clarity/summarization, E-E-A-T signals, Q&A format, and section structure showed positive associations. Their study deliberately excluded schema and technical markup to isolate content signals.

Evergrow Marketing (covered in CMSWire) ran a controlled local-business experiment with LocalBusiness schema. Traditional search rankings (Google, Bing, Maps, etc.) showed no meaningful lift. One exception appeared in ChatGPT for local recommendations, where schema produced a measurable positioning improvement.

What We Saw on Flozi's Own Blog

We looked at how our own blog posts performed in Google Search, split by whether they carry schema markup. We scanned the structured data on every live blog post on flozi.io. We then pulled Google Search Console performance for the same 12 weeks (July 5 to September 26, 2026). Twelve posts have schema (mostly FAQPage, HowTo, and BlogPosting). Twenty-one do not.

What stood out.

• The best performer has no schema. Our IndexNow setup guide earned 3,235 impressions and 33 clicks with no structured data at all. It accounts for most of the no schema group's totals. Without it, that group drops to 1,708 impressions and 4 clicks.
• Schema posts ranked higher on median, but that proves little. The posts with schema are older and sit closer to our core topics. Age and topic are the same confounders behind the correlation in the Ahrefs data.
• The volume is too low to read. The groups differ by a handful of clicks. At this size, one page moves the whole result.
  
  What we can and can't say. Posts with schema showed no clear advantage in search performance, and our top page by clicks has none. That fits the findings above. We can't speak to AI citations, because Search Console doesn't report them and we don't track them. We also can't say whether adding schema to a post that lacked it would change anything.

Google’s Official Position

Google has repeatedly stated that structured data isn't a ranking factor and isn't required for its AI features. John Mueller reiterated that “Structured data won’t make your site rank better.” It is primarily used to enable specific search features and rich results listed in Google’s documentation.

Google’s own guidance on generative AI features explicitly says there is “no special schema.org structured data that you need to add” and lists over-focusing on structured data under myths. Schema remains useful for traditional rich results but is not positioned as an AI visibility lever.

Google has also deprecated several structured data types over time (including most FAQ rich results), further indicating reduced reliance on publisher-provided markup as its systems improve at understanding content directly.

Where Schema Still Matters

Schema is not useless. It continues to power valuable rich results for:

• Products (prices, availability, ratings)
• Recipes
• Events
• Job postings
• Local businesses (in certain surfaces)
• Videos, breadcrumbs, and other supported types

These can improve click-through rates in traditional search. Unused or mismatched markup is generally harmless (Google has said so), but adding types with no live rich-result support creates maintenance cost with little return. Markup must accurately reflect the visible page content.

Practical Recommendations for SEO, AEO, and GEO

1. Do not treat schema as a primary AI citation tactic. The controlled evidence does not support it for pages already in the consideration set.
2. Prioritize fundamentals that consistently correlate with AI visibility: content depth and clarity, E-E-A-T signals, freshness, authoritative backlinks, brand mentions on platforms like Reddit and Quora, and strong domain-level trust.
3. Keep schema for rich results where it still delivers measurable value.
4. Test on your own site if possible, use matched control pages and track AI citations before/after changes.
5. Focus on visible content. If a fact matters for retrieval or citation, state it clearly in the prose AI systems actually read.

The data make one thing clear: schema markup remains a useful technical tool for traditional search features, but the idea that it reliably drives AI citations or rankings has been overstated. Strong content, authority, and user value continue to matter far more.