Short answer. Cited pages carry JSON-LD schema roughly three times more often than uncited pages — but when Ahrefs added schema to 1,885 pages against 4,000 matched controls, the change in citations was statistically indistinguishable from zero on all three engines tested. Both results are correct. Schema is infrastructure: it removes parsing ambiguity, anchors entity identity and qualifies you for features that have documented rules. The genuinely under-invested layer is entity identity — whether a machine can resolve who you are at all.
Structured data is the most over-sold item in the AEO toolkit. It is sold as a citation lever, which the controlled evidence does not support, and dismissed as pointless, which the mechanism does not support either. This piece puts the correlation and the controlled test side by side, reconciles them, and separates what has evidence behind it from what is repeated because everyone repeats it.
What do the two studies actually show?
This is one of the few places in AEO where a correlation and a controlled test both exist, and they disagree in the way correlations and controlled tests usually do.
| Study | Design | Result |
|---|---|---|
| Ahrefs controlled schema test, reported by ROI and Shine | JSON-LD added to 1,885 pages, measured against 4,000 matched control pages | Google AI Overviews −4.6%, Google AI Mode +2.4%, ChatGPT +2.2% — all statistically indistinguishable from zero |
| Ahrefs correlational scan | Six million URLs | AI-cited pages carried JSON-LD roughly three times more often than uncited pages |
The reconciliation is not complicated. Pages that carry schema are, on average, pages maintained by organisations that also do everything else properly: technically sound, updated, structured, written by someone who cares. Schema is a marker of that population, not the cause of the citation.
Schema tells you a page was built carefully. It does not make a page worth quoting.
What is schema actually for?
Discarding it is the opposite error. Structured data does three jobs, none of which is "lift citations".
- It removes parsing ambiguity. A price, a date, a review count or an author is stated as a typed value rather than inferred from prose, and inference is where machines get things wrong.
- It anchors identity.
Organization,PersonandsameAsare how a machine establishes that the entity on this page is the same entity as one in an external reference. That is a different problem from ranking, and content quality does not solve it. - It qualifies you for features that have rules. Rich results, knowledge panels and similar surfaces have documented requirements. Those are deterministic. AI citation is not — Google's own structured data guidance is explicit that correct markup does not guarantee a rich result, and its generative AI guidance states that no special AI-only markup is required to appear in generative features.
Treat it as low-cost infrastructure with a real but narrow benefit and the decision becomes easy.
Why is entity identity the layer that matters more?
If schema is over-sold, entity resolution is under-sold, because its failure looks like a content problem.
An answer engine cannot cite a brand it cannot resolve. When "Acme" could be three companies, a product line and a town, the safe behaviour for a retrieval system is to reach for a source it can resolve instead. No amount of publishing fixes that, because publishing more from an ambiguous entity adds noise to the ambiguity.
The mechanics are unglamorous and mostly one-time.
| Signal | What it does | Cost |
|---|---|---|
| One canonical entity home | Gives the entity a single authoritative URL to resolve to | Low |
Organization schema with a stable @id |
Lets every page refer to one entity rather than many | Low |
sameAs to external identifiers |
Connects your entity to records the engines already resolve | Low |
| A Wikidata item with a stable identifier | Supplies a machine-readable identifier the graph already uses | Low, editorially governed |
| Consistent naming everywhere | Prevents one organisation reading as several | Ongoing discipline |
| Named authors with real, checkable identities | Attaches expertise to a resolvable person, not a byline | Ongoing |
Consistency is the part that decays. Keep the official name, service descriptions, locations, leadership details and headline claims aligned across the site and the important external profiles; when facts conflict, the systems trying to connect them guess worse. Named authorship works the same way — a byline connecting to a real bio, without inflated credentials, and where no individual can be named, transparency about the editorial owner rather than an invented persona.
Wikidata and Wikipedia are community-governed with notability standards and conflict-of-interest rules. Creating or editing entries about your own organisation without disclosure breaks those rules and is routinely reverted. The legitimate route is ensuring the external references those communities require actually exist. That matters commercially because of where Wikipedia sits in the citation graph: the AI Platform Citation Source Index 2026, a synthesis of six studies covering more than 680 million citations, puts Reddit first at roughly 40% of aggregate multi-engine citation frequency and Wikipedia second, appearing in 26–48% of ChatGPT top-10 answers.
What does the evidence say does move citations?
For contrast, here is the intervention with a controlled result behind it. Aggarwal et al., "GEO: Generative Engine Optimization" (ACM SIGKDD 2024), benchmarked content modifications across roughly 10,000 queries and nine datasets: targeted content changes raised visibility in generative engine responses by up to 40%, authority-style edits — adding citations, statistics and quotations — outperformed cosmetic edits such as rewriting, simplification and keyword work, and keyword stuffing performed worse than making no change at all. Effect size varied by domain. We cover that study in detail in what gets you cited by AI answer engines.
Position on the page matters alongside it: Omnibound's 2026 AEO statistics compilation found 55% of sampled AI Overview citations came from the first 30% of the cited page.
Adding a sourced statistic to a paragraph has an evidence base. Adding FAQPage markup to that same paragraph does not. Both take about the same amount of time, and the allocation follows from that.
Strong evidence here means primary research, official documentation, first-party measurement, transparent methodology and clearly sourced statistics — placed next to the claim they support, not collected in a footer. It includes limitations: a page that explains where a method fails is more credible than one claiming universal success.
The cargo-cult list
Practices that recur in AEO proposals and have no supporting evidence behind them.
- Adding FAQ schema to lift citations. Correlational at best; the controlled test found nothing. Use it only where the visible page genuinely contains the questions and the site meets current eligibility rules.
- Publishing llms.txt as a visibility tactic. Google's documentation states that Search does not use llms.txt or similar AI text files, no major crawler claims to read it, and adoption studies reported by Digital Applied found 97% of published files received zero traffic across 137,000 sites. Harmless to publish, misleading to count — and it routinely displaces the crawler-access check, which genuinely does decide whether you can be cited.
- Marking up every entity type available. Schema you cannot maintain is schema that will eventually contradict the page.
- Schema that does not match the visible page. A mismatched
FAQPageblock is ignored at best and a quality problem at worst. Unsupported review ratings and fabricated author credentials are policy exposure, not optimisation. - Treating markup coverage as a KPI. It measures effort, not outcome.
- Manufactured third-party mentions. Bought reviews, placed listicles and citation farms create short-term mentions and weaken the trust they were meant to build.
What does a defensible policy look like?
- Implement the identity layer once and maintain it.
Organization, a stable@id,sameAs, consistent naming, named authors. This is the part with a real mechanism behind it. - Implement
Article,BreadcrumbListandFAQPageonly where they match the page exactly. Cheap hygiene, no claimed lift. - Generate markup from the same source of truth that renders the page, so titles, authors, dates and images cannot drift apart, and validate on every publish.
- Stop reporting markup coverage as an AI visibility metric. Report citation rates instead.
- Spend the freed effort on sourced specificity, front-loaded. That is the intervention with a controlled result behind it.
- Monitor whether AI systems describe you accurately. A wrong description signals that the web evidence is incomplete, inconsistent, or outweighed by a stronger third-party source.
Three honest limits. One controlled study is one controlled study — the Ahrefs test measured adding schema to pages that were already visible, and does not rule out an effect on pages with no other signals. Requirements for deterministic features are real and separate, so losing a rich result because markup was removed is a genuine cost unrelated to AI citation. And entity resolution has no published effect size: the mechanism is well understood, a controlled quantification is not, and this article does not claim one.
How Lifewood approaches this
Lifewood implements the identity layer as a one-time engineering task and then treats markup as hygiene rather than a line item with a projected citation lift attached. Schema is generated from the same source of truth that renders the page and checked against the visible text on every publish, because markup that contradicts the copy is worse than none.
The effort freed by not chasing markup coverage goes into sourced specificity and into the entity work that has a mechanism: one canonical home, consistent naming, external references that resolve. Across 50+ languages and 40+ delivery centres across 30+ countries, naming consistency is the hardest part to hold, because an organisation described three ways in three markets reads as three entities.
See AEO services, GEO services, AEO and GEO providers and the glossary.
Sources and further reading
- Ahrefs controlled schema test (1,885 pages, 4,000 matched controls) and six-million-URL correlational scan, reported by ROI and Shine, 2026.
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, GEO: Generative Engine Optimization, ACM SIGKDD 2024.
- Omnibound, Answer Engine Optimization statistics 2026 — position-in-page figures.
- AI Platform Citation Source Index 2026 — synthesis of six studies covering more than 680 million citations.
- Google Search Central — general structured data guidelines, and guidance on optimising for generative AI features.
- SE Ranking and Ahrefs llms.txt adoption studies, reported by Digital Applied, 2026.