Adding structured data to your site to improve AI search visibility sounds like a safe bet. It is not. A February 2026 empirical study of 730 AI citations found that generic, partially-filled schema produces an 18-percentage-point citation penalty compared to having no schema at all. Meanwhile, a May 2026 Ahrefs controlled experiment across 1,885 pages showed that adding JSON-LD produced no statistically significant citation increase on ChatGPT or Google AI Mode. The picture is more complicated — and more actionable — than the “just add schema” advice circulating in GEO guides.
Here is what the data actually shows, and the specific implementations that have measurable impact versus the ones that are noise.
Why Generic Schema Hurts More Than No Schema
The 18-point citation penalty from partial schema is the most important finding to understand first, because it runs counter to most advice. The study examined 730 citations generated by ChatGPT (GPT-4o with web browsing) and Gemini (1.5 Pro with search grounding) across 75 commercial queries in five categories: SaaS and Technology, Health and Medical, Finance and Insurance, Professional Services, and How-To and DIY.
Pages with no schema at all outperformed pages with generic or incomplete structured data. The researchers concluded that AI systems treat partially-filled schema as a negative signal — evidence that a page attempted to use schema but did not maintain it accurately. An Organization schema with an empty url field, an Article schema with a placeholder author value, or an FAQPage schema where the answers duplicate the page headings verbatim all scored worse than pages that had no schema at all.
The implication is direct: if your schema is not accurate and complete, removing it may improve your AI citation rate. The bar is higher than most implementations meet.
The Ahrefs Controlled Study: What Adding JSON-LD Actually Did
The May 2026 Ahrefs experiment provides the cleanest data available. The team identified 1,885 pages that added JSON-LD structured data between August 2025 and March 2026, then matched them against 4,000 control pages with similar pre-treatment citation rates. Citation rates were measured 30 days before and 30 days after schema addition.
Results by platform:
- Google AI Overviews: down 4.6% — statistically significant negative result
- Google AI Mode: up 2.4% — not statistically significant
- ChatGPT: up 2.2% — not statistically significant
The 4.6% drop in AI Overviews citations is notable and likely reflects the same dynamic as the earlier study: sites adding schema often do so inconsistently, and the inconsistency itself is penalized. The null result on ChatGPT and AI Mode means that schema alone — without the content quality signals AI systems actually weight — does not move the needle.
What Actually Works: Entity Disambiguation Schema
Despite the mixed aggregate results, one category of schema implementation shows consistent positive correlation with AI citation rates: entity disambiguation. This is Organization schema that uses sameAs to link your brand to external authoritative identifiers — Wikidata, LinkedIn, Crunchbase, GRID (for research organizations), and industry directories.
The mechanism is different from traditional SEO. AI systems built on large language models are trained on documents that refer to entities by multiple names. Without disambiguation, your brand may be treated as an ambiguous reference rather than a known entity. When your Organization schema explicitly links your domain to a Wikidata entity ID, LinkedIn company page, and Crunchbase profile, AI models can anchor your brand to their existing knowledge about it — and cite it with higher confidence in contexts where that entity is relevant.
A complete entity disambiguation implementation looks like this:
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Your Company Name",
"url": "https://yoursite.com",
"logo": "https://yoursite.com/logo.png",
"sameAs": [
"https://www.wikidata.org/wiki/Q[YOUR-ID]",
"https://www.linkedin.com/company/[YOUR-SLUG]",
"https://www.crunchbase.com/organization/[YOUR-SLUG]"
],
"knowsAbout": [
"AI search visibility",
"generative engine optimization",
"your primary topic area"
]
}
</script>If your brand does not have a Wikidata entry, creating one is worth the 20 minutes it takes. Wikidata is one of the highest-confidence entity resolution sources used by AI systems trained on Wikipedia and its associated datasets.
The Four Schema Types With Measurable AI Citation Impact
Outside entity disambiguation, four schema types show consistent correlation with AI citation across multiple platform studies:
FAQPage — conditional on implementation quality. Pages with FAQPage schema where the questions match actual user search queries and the answers are 40-120 words of original content outperform both pages with no schema and pages with FAQPage schema that uses keyword-stuffed questions or very short answers. The critical variable is whether the questions reflect how users actually ask about your topic — not how you prefer to frame it. Use Search Console query data to identify real questions before writing FAQ answers.
HowTo — effective for procedural content. HowTo schema lets AI systems extract your process as a structured sequence rather than prose. Pages using HowTo schema for genuinely procedural content (not retrofitted onto opinion pieces) see higher citation rates in “how to” query responses across Perplexity and ChatGPT. Each HowToStep should be a discrete action, not a section heading.
Article with complete author entity. The author field in Article schema should reference a Person entity with a complete profile — name, url pointing to an author bio page, and if possible a sameAs link to the author’s LinkedIn or Google Scholar profile. Article schema where the author field is “Content Team” or left empty is worse than no Article schema at all, per the February 2026 study.
Speakable — underused and underrated. Speakable schema explicitly marks sections of your content as answer candidates for voice and AI responses. Despite being a Google-specific implementation, its use has expanded to ChatGPT citation patterns as well. Mark your clearest, most direct answers — typically the first paragraph after each subheading — as speakable content. This signals to AI parsers exactly which text is intended as a citable summary.
The 41 Percent Adoption Problem
JSON-LD is now on 41 percent of all web pages, up from 34 percent two years ago per the HTTP Archive Web Almanac structured data chapter. The problem is that adoption outpaced implementation quality. Most of that 41 percent is boilerplate schema generated by WordPress plugins with default fields, e-commerce platforms with autogenerated Product schema, and CMS templates that add Organization schema once at install and never update it.
This creates an opportunity. When 41 percent of the web has schema and most of it is low-quality or inaccurate, sites that implement schema correctly — complete, accurate, regularly audited, entity-linked — stand out. The signal value of good schema is higher precisely because bad schema is so common.
To audit your current schema: use Google’s Rich Results Test on your key pages, check that every required field is populated with accurate data (not placeholder text), and verify that entity links in sameAs fields resolve to live, correct pages. Run this audit quarterly — company URLs change, LinkedIn slugs get updated, and stale schema is worse than none.
Implementation Priority Order
If you are starting from zero or auditing an existing schema setup, work in this order:
- Organization schema with sameAs entity links — applies site-wide, highest leverage for entity recognition across all AI platforms
- Article schema with complete Person author entity — on every piece of content that targets informational queries
- FAQPage schema — only on pages with genuine Q&A sections where answers are 40+ words of original content
- Speakable markup — on your clearest answer paragraphs, particularly on product, service, and FAQ pages
- HowTo schema — on procedural guides with discrete steps
Do not add schema to every page type at once. Implement one type, validate it fully, then move to the next. Inconsistent schema across your site is worse than consistent schema on a subset of pages.
What Schema Cannot Fix
Structured data is a signal layer on top of content. It helps AI systems extract and attribute information from pages that already contain credible, citable content. It does not compensate for thin content, unnamed authors, lack of original data, or a site with no external references pointing to it.
The Ahrefs study’s null result on ChatGPT is instructive here: ChatGPT citations are driven more heavily by training data and real-time web browsing of credible pages than by schema parsing. Adding JSON-LD to a page that is not otherwise being cited because of its content quality will not produce citations. Schema improvements work on the margin — they help AI systems correctly attribute content that is already being considered, not push mediocre content into the citation set.
If your citation rate is near zero, the problem is content and authority, not schema. Fix those first, then layer in structured data to capture the full value of the content you have built.
Conclusion: Audit Before You Add
The 2026 data on schema markup delivers a clear message: inaccurate or generic structured data is worse than none, entity disambiguation schema has the highest consistent impact, and schema alone does not move AI citation rates for content that lacks underlying authority signals.
Before adding more schema, audit what you have. Remove or fix incomplete implementations. Then add entity disambiguation, complete author entities, and FAQPage schema where it genuinely applies. That sequence produces measurable results. Generic plugin-generated schema on every page does not.
Want to see how your site’s structured data looks to AI systems right now? Run a free scan at ai-visibility.llmagnet.com — it checks schema implementation alongside 11 other AI readiness signals and returns results in under 5 minutes.