Does Schema Markup Help AI Citations? What the Evidence Actually Shows

Barely. The best causal study to date — Ahrefs, 1,885 pages that added schema against roughly 4,000 matched controls — found AI citations did not rise, and slipped 4.6% in AI Overviews. Google states structured data is not required for its AI features. Add schema because it is free and helps classic Search, not because someone sold it as the unlock.

A roofing contractor pays an agency $2,400 for an “AI search readiness” package. The deliverable arrives 6 days later: structured data across 40 pages, a validator screenshot with 0 errors, and a report calling the site AI-ready. Four months on, asking ChatGPT who to hire for roof replacement in that metro still returns 3 competitors and no mention of the client.

Nothing was broken. The markup is valid, the pages render, the validator is genuinely green. The work simply had almost nothing to do with the outcome it was sold against — and the reason is measurable rather than a matter of opinion.

Does schema markup help AI citations?

Not meaningfully. Ahrefs tracked 1,885 pages that added structured data against roughly 4,000 matched control pages and found citations barely moved — down 4.6% in AI Overviews. Google states structured data is not required for its AI features. Schema earns rich results in classic Search; it does not buy a mention.

The study design that makes it credible

Most claims about AI search are correlational: someone notices that cited pages tend to carry structured data and concludes the markup caused the citation. That inference fails because pages carrying good schema also tend to be on larger, better-maintained, more-linked sites — the markup is a symptom of a competent site, not the cause of the citation.

The Ahrefs work matters because it used matched controls. Pages that added schema were compared against roughly 4,000 comparable pages that did not, which isolates the change from everything else that differs between sites. That design is what turns “cited pages have schema” into “adding schema did not produce citations.”

A 4.6% decline is best read as noise around 0 rather than as evidence that schema hurts. The honest conclusion is not that structured data is harmful. It is that the effect size is indistinguishable from nothing.

Why do agencies still sell schema as the AI unlock?

Because it is the most sellable deliverable in the category. Schema is a one-time technical change, shippable in under 1 day, and it produces a green validator screenshot that reads as proof. The 4 levers that actually move citations are slower, need editorial judgement, and resist being packaged into a fixed-scope invoice.

The incentive, traced

Consider what each option looks like on a proposal:

• Schema markup — fixed scope, 1 developer, delivered in days, verifiable with a free validator.
• Answer restructuring — requires rewriting pages someone already approved, and the client must review every change.
• Entity consistency — needs access to profiles the agency does not control, and 3 rounds of the client finding old listings.
• Third-party corroboration — cannot be guaranteed at all, because it depends on other people publishing.

Only the first can be promised with a date. That is the whole mechanism. It is not usually dishonesty so much as an agency shaping its offer around what it can commit to, and a buyer preferring the option with a delivery date attached.

The tell is the pitch itself. Anyone selling structured data as the thing that gets you into ChatGPT has not read the evidence — a position DaxReach states publicly on its comparison page and in the GAS Engine Playbook.

What actually moves AI citations?

Four levers, in strict dependency order: crawler access, because nothing downstream matters if the page cannot be read; extractable answers of roughly 40 to 60 words under a question-shaped heading; entity consistency across every profile an engine can find; and third-party corroboration from sources you do not own.

The 4 levers in detail

1. Crawler access. Check robots.txt for the 9 user agents that matter — OAI-SearchBot, ChatGPT-User, GPTBot, PerplexityBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot and Bytespider. A site that disallowed these in 2023 to stay out of training data also removed itself from recommendation. This is usually a 1-line fix.

2. Extractable answers. An engine lifts a self-contained claim, not a page. Answers running 150 words past the heading get truncated or skipped in favour of a competitor who said the same thing in 50. The mechanics are covered in what answer engine optimization is.

3. Entity consistency. Name, category, address and phone must match across your site and every profile. Where they conflict, an engine either picks one or declines to name you. This is unglamorous and it is where most local businesses lose.

4. Third-party corroboration. What independent sources say about you carries more weight than what you say about yourself, which is precisely why it cannot be bought as a deliverable. It accrues.

Schema sits below all 4. It is worth having and it is not a lever.

Which lever should you buy first?

Whichever one is currently failing, tested top-down and stopped at the first failure. There is no value in rewriting answer blocks on a site whose crawlers are disallowed, and no value in chasing citations while 2 profiles disagree about your business category. Diagnose in dependency order, then spend.

The levers, side by side

Read the “weakest at” column honestly: crawler access is nearly free and often already fine, so it is not a project. Extractable answers are the usual best first spend for a site with existing rankings, and useless for a site with 4 thin pages.

The 3 routes to getting it done

The in-house column is not a strawman. A business with an in-house marketer who reads the AEO versus SEO breakdown and works the 4 levers in order will beat any outsourced arrangement, because they know the market better than a vendor can. The constraint is almost never capability. It is whether the work survives contact with a busy quarter.

Should you add schema anyway?

Yes — just stop paying a premium for it. Structured data still drives rich results in classic Google Search, most content systems emit it automatically, and the marginal cost after the template is written is 0. Keep it, confirm it parses, and move the remaining budget to access, answers, entities and corroboration.

The classic-Search case for keeping it

Schema is infrastructure, and infrastructure is judged by whether it works, not by whether it is exciting. Review stars, breadcrumbs and product details in ordinary blue-link results still influence click-through on the traffic that pays the bills today. That case stands entirely on its own merits and needs no generative story attached to it.

What should change is the budget line it sits on. Schema belongs in technical SEO, priced accordingly, next to sitemaps and canonical tags. It does not belong on an AI-visibility invoice, and a vendor who puts it there is either not reading the evidence or is counting on you not to.

The full vocabulary is in the glossary, the scoring method behind these signals is in what the GAS score measures, and how this work is actually run is on how it works and AI Visibility.

Schema markup is a reasonable thing to have and a poor thing to buy. The 4 levers above are the ones that decide whether an engine says your name.