In May 2026, Ahrefs published the most rigorous independent study of structured data and AI citations to date. They tracked 1,885 web pages that added JSON-LD schema between August 2025 and March 2026, matched them against 4,000 control pages, and measured what happened to their citation rates across Google AI Overviews, Google AI Mode, and ChatGPT.
The result contradicts two years of GEO advice: adding schema produced no statistically meaningful citation uplift on any major platform. Google AI Mode increased by 2.4%, ChatGPT by 2.2% — both within the margin of error. Google AI Overviews actually dropped 4.6% in the schema-added group.
This matters because “add schema markup” has become the default first recommendation in almost every GEO guide published since 2024. If schema doesn’t move the needle on citations, what does? Here’s what the data actually shows.
Why Schema Doesn’t Work the Way GEO Guides Claim
The intuition behind schema advice is sound: if you label your content in a machine-readable format, AI retrieval systems should be able to parse and cite it more reliably. The problem is that AI citation doesn’t work through schema parsing — it works through retrieval and relevance scoring.
When an AI platform like ChatGPT or Perplexity generates a response using retrieval-augmented generation (RAG), it retrieves pages based on semantic relevance to the query, then evaluates whether the retrieved content contains specific, grounded, factually useful claims. JSON-LD schema isn’t part of that pipeline. The model doesn’t read your FAQPage markup and decide to cite you — it reads your actual text and evaluates whether it directly answers the question.
The Ahrefs study’s control group confirms this. Sites that added schema tended to also invest in technical SEO, authoritative content, and link building. When you strip those confounders out by matching schema pages against comparable control pages, the schema signal disappears. The citation advantage was coming from the content, not the markup.
What the Princeton Study Found Instead
The Princeton + Georgia Tech + IIT Delhi GEO study — the most-cited academic paper on AI search optimization — tested nine content modification strategies on real AI platforms and measured which ones actually increased a page’s visibility in AI-generated answers.
The top performers were not technical: they were textual. Adding direct quotations from authoritative sources produced a 41% visibility increase. Adding specific statistics lifted visibility by 32%. Explicitly citing your sources in-text (not via schema — just attributing claims to named sources) produced a 30% improvement. Adding fluency and authority language (“according to,” “research shows”) delivered 28%.
None of the top four tactics require touching your site’s code. All four require changing what you write and how you attribute it. The implication is that AI retrieval rewards content that sounds like it comes from a credible source — not content that is technically marked up as one.
Platform Divergence: One Strategy Doesn’t Cover All AI Search
A 2026 analysis by Averi.ai covering B2B SaaS citation patterns across ChatGPT, Perplexity, and Google AI Mode found that only 11% of domains are cited by both ChatGPT and Perplexity. Google AI Overviews and Google AI Mode share only 13.7% of their cited URLs. These platforms are not drawing from the same source pool, and a single optimization strategy won’t cover all of them.
The divergence runs deep:
- ChatGPT leans heavily on Wikipedia-style encyclopedic content (47.9% of its top citations) and favors longer-form pages, citing content that averages 393–458 days newer than typical Google results
- Perplexity cites Reddit in 46.7% of its responses, cites 21.87 sources per answer (vs. ChatGPT’s 7.92), and strongly favors structured H2/H3 pages organized around specific questions with visible data and named methodology
- Google AI Overviews still overlaps heavily with traditional search — 40.58% of its citations come from Google’s top 10 organic results — and shows almost no citation preference for longer content
The practical implication: if you optimize only for schema and technical signals, you’re building a strategy that none of these platforms primarily respond to. If you optimize for content structure, specificity, and attribution — you cover all three.
The Four Content Changes That Actually Drive AI Citations
Based on the Princeton study data, the Averi.ai platform analysis, and the Ahrefs schema study, here are the four content changes with the strongest citation evidence behind them:
1. Add attributed statistics to every key claim. Replacing “AI search is growing rapidly” with “AI-generated answers now influence 40% of B2B vendor research according to Gartner’s Q1 2026 survey” gives the AI retrieval system something specific and verifiable to cite. Unattributed claims are hard to ground; named, dated statistics are easy. The Princeton study found a 32% visibility increase from statistics addition alone.
2. Include direct quotations from identifiable sources. This is the single highest-impact tactic in the Princeton study (+41%). Quote researchers, executives, or published reports by name. “According to Dr. Chirag Shah at the University of Washington” gives an attribution chain that AI systems can verify and follow. Generic summaries give them nothing.
3. Structure pages around specific answerable questions. Perplexity’s citation behavior strongly favors pages where individual H2 or H3 sections directly answer a discrete question. Instead of “Our Approach to AI Visibility,” write “How Does AI Visibility Differ From Traditional SEO?” Each section should be independently citable — meaning a reader could extract it, and it would fully answer the question it’s addressing without the surrounding context.
4. Target the platforms where your audience actually searches. If your buyers are B2B decision-makers, ChatGPT and Perplexity are more likely touchpoints than Google AI Overviews. For ChatGPT, that means encyclopedic depth and authoritative sources. For Perplexity, it means community-validated content and specific data. Optimize for both by writing content that is simultaneously detailed and directly question-answering.
What Schema Is Actually Good For
The Ahrefs finding doesn’t mean schema markup is worthless — it means its value is in the wrong place for GEO purposes. JSON-LD schema still matters for:
- Google’s traditional rich results (star ratings, FAQs, how-tos in standard SERPs)
- Entity disambiguation — consistent
OrganizationandProductschema helps search systems correctly identify what your brand is and what it offers, which can improve indirect citation through better entity matching - Breadcrumb and site structure signals that help crawlers understand your content hierarchy
Add schema for these reasons. Don’t add it expecting AI citation lift — the controlled data shows it doesn’t deliver that.
The Checklist: What to Change Before Your Next Content Audit
If you’re running a GEO content audit, this is what the evidence supports checking:
- Statistics sourcing — every data point should have a named source and a date. Replace “studies show” with the actual study name.
- Quote density — aim for at least one direct quotation from a named expert or report per 400 words.
- Section structure — each H2 should be answerable as a standalone question. Rewrite section headers that describe topics (“Our Methodology”) as questions (“How Does Our Methodology Work?”).
- Freshness — Ahrefs’ 17 million citation study found AI platforms cite content 25.7% fresher than traditional search. Statistics older than 18 months in fast-moving fields need updating.
- Entity consistency — your brand name, product name, and key terminology should be identical across every page, meta description, and structured data field. Variation confuses entity resolution.
None of these changes require a developer. All of them have stronger citation evidence behind them than schema markup.
Conclusion
The schema advice spread through GEO content because it sounds plausible and it’s technically implementable. The Ahrefs study shows it doesn’t deliver the citation impact that’s been attributed to it. The content-level changes — statistics, quotations, question-structured sections, current data — consistently outperform technical markup in actual citation measurement.
The most impactful GEO investment you can make right now is a content audit that adds attribution to your existing statistics, restructures your section headers as answerable questions, and updates any data that’s more than 18 months old. That work compounds directly into citation rate; adding JSON-LD to the same pages doesn’t.
Want to see how your current pages score across ChatGPT, Perplexity, and Google AI Overviews — and which specific signals are missing? Run a free audit at ai-visibility.llmagnet.com. It checks 47 visibility signals in under 60 seconds and flags the exact gaps holding your citation rate down.