Does Schema Markup Help AI Citations? What the Evidence Shows

Schema markup is the most oversold tactic in AI search. The pitch writes itself — "machines read structured data, LLMs are machines, therefore mark everything up" — and an industry of implementation retainers hangs off that syllogism. The evidence does not cooperate: the one controlled study on the question found nothing, and Google's own documentation disclaims any markup requirement for its AI features. This page lays out what the evidence supports, what it doesn't, and why we still ship JSON-LD on every page we publish.

Does adding schema get you cited by AI engines?

No — the only controlled test of the question found no causal lift. Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matched them against roughly 4,000 control pages that didn't, and ran a difference-in-differences analysis with 4 separate statistical tests. The conclusion, published May 2026: "adding schema produced no major uplift in citations on any platform."

The per-platform numbers deserve quoting because they are so unlike the marketing claims built on them:

Ahrefs, 1,885 treated pages vs ~4,000 controls, August 2025–March 2026.

A −4.6% on AI Overviews and two zeros is a null result with a slight negative edge — the opposite of a tactic. The study's scope limits matter too: it measured pages already visible to AI systems, and its author notes schema may still play a role in how undiscovered pages get found. That is an open question, honestly labeled, not evidence for the tactic.

Why do 53% of AI-cited pages have schema, then?

Because correlation is doing all the work. Ahrefs' starting observation was that 53% of AI-cited pages carry schema — the number vendors quote. But sites that invest in technical SEO tend to add schema and build crawlable structures, fast pages, and linkable authority at the same time. As the study put it, schema "could be doing real work, but it could also just be riding the wave of every other signal."

The controlled comparison exists precisely to separate the rider from the wave, and when it did, the wave was everything: pages that added the markup gained roughly nothing over identical pages that didn't. This is the correlation-vs-causation trap this niche falls into constantly — the same pattern behind most "we analyzed 10,000 citations and found X" posts, where X describes what cited sites have, not what citing engines reward. How engines actually select sources is a retrieval-and-passage story, walked through in how AI search works.

What does Google say about schema and AI features?

That none is required — in unusually plain language. Google's AI-features documentation: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." Eligibility for AI Overviews and AI Mode is 1 rule: indexed, and eligible to show in Google Search with a snippet.

Note what Google is not saying: that structured data is useless. Its structured-data documentation still describes markup as helping Google understand page content and as the eligibility mechanism for rich result features — and its guidelines require markup to reflect the visible content of the page rather than claims the reader can't see. The documented position, held together: schema feeds specific Search features; AI features carry no schema requirement of their own.

The Princeton GEO paper points the same direction from the research side: the tactics that moved visibility in its tests — measured at up to 40% — were content-level changes like adding citations, quotations, and statistics to the text itself. The wins were in the words, not the markup.

Where does schema still earn its keep?

Two documented jobs survive the null result untouched, and they are why we keep shipping it. First, rich results: Google's structured-data system is the entry ticket to result enhancements, which exist regardless of what AI engines do. Second, entity disambiguation: Organization and similar markup states, in machine-readable form, who published a page and how that publisher relates to its other properties — plumbing that helps any system trying to resolve who is asserting this.

Across our fleet — 400+ published pages on 3 production builds — every page carries JSON-LD appropriate to its type, and we maintain it for those 2 reasons, not for citation lift [our data]. Worth stating plainly: we run this markup in production and still tell you it will not earn citations, because our test of record is Ahrefs', and it found nothing.

One more layer, labeled at its evidence tier: whether AI assistants even read JSON-LD at answer time is undocumented — no platform on our source list states that it does — while what engines demonstrably quote is visible passage text. Treat "LLMs parse your schema when answering" as plausible at best and unsupported in the documentation. If you want a lever the quoting behavior clearly runs through, structure the visible words: that is the citable passage — the self-contained, quotable block — and it is where the Princeton results actually point.

Should you pay for schema implementation?

Not on an AI-citations pitch — the evidence is now specific enough to disqualify that sale. A retainer justified by "structured data gets you into AI Overviews" is contradicted by Google's documentation and by the only controlled study run on the question. If a vendor quotes the 53% figure at you, ask whether they can explain why the controlled version of that number is roughly 0.

What is worth doing: basic, honest markup — Article or its equivalent per page, Organization once, kept consistent with visible content — implemented once and templated. It is cheap, it serves the documented jobs above, and on most stacks it is an afternoon of work rather than a monthly line item. Then spend the freed budget where the evidence says citation selection actually happens: answer-first pages, extractable passages, and claims specific enough to be worth attributing — the practice documented across our generative engine optimization guide and applied to Google's surfaces in how to show up in AI Overviews.