"Add schema markup to get cited by AI" is one of the most-invoiced claims in GEO. In May 2026, it finally met a controlled study — and lost.
The evidence
Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matched against ~4,000 control pages with similar prior citation levels, and measured AI citations 30 days either side. The result, in their own words: adding schema produced no major uplift in citations on any platform — Google AI Overviews, AI Mode or ChatGPT. (The one statistically significant movement was a small decline on AI Overviews.)
A separate experiment they cite (searchVIU) tested what AI systems actually read when they fetch a page live: every system extracted only the visible HTML content. JSON-LD and hidden markup were ignored at retrieval time.
Yes — pages that get cited are more likely to have schema. That's correlation: well-run sites do both. When you isolate the variable, the citation effect disappears.
So is schema dead? No — it was mis-sold
We updated our audit's structured-data guidance (changelog) to score schema for what the evidence supports:
- Entity clarity. Organization, Person,
sameAs,knowsAbout— this is how machines resolve WHO you are and connect you to your real-world profiles. Entity confusion is a genuine, observed failure mode for brands in AI answers; schema is the cheapest tool against it. - Microsoft surfaces. Bing states plainly that structured data helps its systems — and Bing's index feeds Microsoft Copilot. One platform, confirmed by the platform itself.
- Machine-readability hygiene. Prices, availability, authorship, dates in parseable form get misread less — by anything consuming your site.
What we no longer do: score schema as a citation booster, recommend HowTo markup (rich results removed in 2023), or promise FAQ rich results (Google deprecated them for everyone in May 2026 — though visible Q&A content remains one of the strongest extraction patterns there is).
The practical takeaway
Spend your effort in this order: visible, extractable content first — the answers, definitions and statistics engines actually lift; entity schema second — Organization/Person with real sameAs profiles; decorative markup last, if at all. That's the order our audit weights it, and every report states which methodology version scored it — so when the evidence moves again, you'll see exactly what we changed.