Schema Markup Doesn't Drive AI Citations. It Still Matters.

Most claims in this space are correlational: pages with schema get cited more, therefore add schema. That reasoning collapses the moment you notice that the kind of site which implements structured data properly is also the kind of site which does everything else properly.

This study avoided that trap. Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, and matched them against roughly 4,000 control pages. The primary analysis was difference-in-differences: rather than comparing pages with schema to pages without, it compares how the treated pages changed against how the controls changed over the same period. That isolates the effect of the change itself.

That's a genuinely good design, and it's rare in this field. They also ran event studies and t-tests as validation, which is more rigour than most vendor research shows.

The authors' own summary is the honest one: adding schema produced no major uplift in citations on any platform, and the most consistent finding is that not much really changed.

Don't over-read the −4.6%. A small negative on one surface, in one study, is not evidence that schema hurts you. The finding is the absence of an effect, and the absence is what's interesting, because an entire category of advice has been built on the assumption that it would be there.

Once you know how retrieval works, the result stops being surprising. It's what Article 3 of this series would have predicted.

RAG retrieval operates on semantic similarity between a query embedding and a passage embedding. The passage is prose. Your JSON-LD block sits in the head of the document, describing the page in a structured vocabulary designed for a different consumer entirely: search engine features that need to know this is a recipe, or a product, or an FAQ, so they can render a specific interface component.

Nothing in that pipeline is waiting for a schema block to tell it what your page means. The model reads the passage. Structured data was built to trigger rich results, and it does that job well. It was never a semantic hint to a retriever, and expecting it to behave like one was a category error.

Schema tells a search feature what to render. It doesn't tell a retriever what your page means.

I'd rather over-correct here than have somebody strip their structured data on the strength of a headline.

Schema still earns rich results in traditional search: FAQ accordions, review stars, recipe cards, breadcrumbs, event listings. Those are real SERP features with real click-through consequences, and they are the job structured data was designed for. Removing it to chase a finding about AI citations would be trading something that demonstrably works for nothing at all.

I'll be concrete about my own practice, because it would be odd to write this and quietly do the opposite. Every article on this site carries Article and FAQPage schema, and I run a linter over the whole blog that fails the build if the visible FAQ and the FAQPage schema fall out of sync. That's not an AI visibility measure. It's rich-result hygiene, and the two things are simply unrelated.

The practical adjustment is smaller than the headline suggests, and it's mostly about honesty in the pitch.

• Stop putting schema in the GEO column. If a proposal lists structured data under AI visibility deliverables, it's now claiming something the best available evidence contradicts. Move it to technical SEO, where it earns its place.
• Stop auditing schema as an AI diagnosis. An audit that reports "no Organization markup" as an AI visibility problem is measuring the wrong layer. The AI visibility problems are access, structure and source-worthiness.
• Spend the effort on passages instead. Retrieval works on prose, so the work that moves AI citations is making each section a clean, self-contained claim. That's unglamorous and it's what the mechanism actually rewards.
• Say it out loud to clients. "We're keeping your schema because it earns rich results, and we're not counting it as AI work" is a sentence that builds trust. It also distinguishes you from whoever billed them for a schema-driven GEO retainer.

There's a pattern running through this part of the series. llms.txt, schema markup, and most of the widely-quoted statistics all share the same shape: an intuitive story, adopted quickly, tested late. When the test finally arrives it usually says less happened than everyone assumed.

The useful habit isn't scepticism for its own sake. It's asking what the mechanism would have to be for a tactic to work, and whether anyone has actually looked.