What actually correlates with AI citation
Key takeaways
- Causal evidence exists for three levers: quotations about +41%, statistics about +34%, cited sources about +29%.
- Citing sources helps challengers most, around +115% for a rank-five page, and slightly hurts a page already ranked first.
- Branded mentions correlate about 0.66 to 0.71 with AI visibility versus about 0.22 for backlinks.
- Adding schema showed no measured citation lift in a matched 1,885-page analysis.
- The circulating multipliers for tables, section length and Person schema trace to no source and were deleted from our catalog.
Three content levers have controlled causal evidence behind them: adding quotations lifted generative visibility about 41%, statistics about 34%, and cited sources about 29% — with the source-citation effect reaching roughly 115% for lower-ranked pages. Almost everything else sold as an AEO tactic is correlational at best, and several widely repeated multipliers trace to nothing at all.
What has actual causal evidence?
The peer-reviewed GEO work presented at KDD 2024 remains the only controlled experimental evidence in this area. It tested content modifications against generative engines and measured visibility change:
| Modification | Measured effect | Note |
|---|---|---|
| Adding quotations | +41% | The largest single content lever tested |
| Adding statistics | +34% | Specific figures, not adjectives |
| Citing sources | +29% | Roughly +115% for pages ranked around fifth; about −30% for pages already ranked first |
| Keyword stuffing | −8 to −10% | Net negative, as it is in classic search |
Note the asymmetry in the third row. Citing sources is a challenger's lever: it helps a page that needs to borrow authority and slightly hurts one that already has it. Any tool that recommends it uniformly has not read the study.
What is strongly correlational?
- Branded web mentions. Correlate roughly 0.66–0.71 with AI visibility across a 75,000-brand sample, against about 0.22 for backlinks. Roughly 82% of citations are earned media. This is the single largest factor in the whole field and it is not on your website.
- Position within the page. 44.2% of ChatGPT citations come from the first 30% of content, measured across 1.2 million answers and 18,000 verified citations.
- Question-form content. Cited roughly twice as often, with 78.4% of question-tied citations attaching to a heading.
- Format matching intent. Listicles take 21.9% of all AI citations and 40.9% of commercial-intent ones; articles take 45.5% of informational citations.
What does not survive checking?
Three things we removed from our own catalog after failing to trace them:
- Schema as a growth lever. A matched difference-in-differences analysis of 1,885 pages that added JSON-LD found no citation lift — slightly negative for AI Overviews, a statistically insignificant positive for ChatGPT. Cited pages are about three times more likely to carry schema, but that is confounded by site quality. Schema is hygiene: keep it valid and non-contradictory, and do not expect it to buy anything.
- llms.txt. Across 137,210 domains, 97% of llms.txt files received zero bot requests, and Google has said it will not support the format.
- The circulating multipliers. "Tables get cited 2.5 times more often", "50 to 150-word sections get 2.3 times the citations", "Person schema is worth 1.8 times". Three independent attempts to source these found nothing. Tables and short sections are still worth building — on mechanism, because they chunk cleanly — but the multipliers are decoration.
Citing sources is a challenger's lever. It helps a page that needs to borrow authority and slightly hurts one that already has it — which is why recommending it uniformly is a sign the study was not read.
— AnswerOpen reading of the GEO results, KDD 2024
What about freshness?
Genuinely contested, and the honest answer is that it depends on the query. One
dataset finds 65% of AI bot hits target content less than a year old and 76% of
ChatGPT's most-cited pages were updated within thirty days. Another, across 1.4
million prompts, finds the median cited page is around 500 days old for evergreen
questions. Both are probably right about different query types. Our position: the
right thing to audit is an honest, machine-visible date triad — a visible updated
date, a matching dateModified, and revalidation headers that agree
with both. We never reward date-churning, because a model that gets caught by a
contradiction between your visible date and your schema has learned something
worse than staleness.
So what would you actually do first?
- Fix access. It is binary and usually cheap.
- Move the direct answer to the top of the page and phrase headings as questions. This is free and touches the strongest correlational finding.
- Add real quotations, real statistics and real cited sources — the three levers with causal evidence — provided they are true.
- Work on branded mentions off-site. Largest factor, slowest lever, and no auditor can do it for you.
- Keep schema valid and consistent. Do not budget for it as growth.
Related: how each of these becomes a check · why order matters. The GEO paper is public: Generative Engine Optimization, KDD 2024.
Primary sources
Everything asserted above traces to one of these. Operator documentation changes often; check the current version before relying on any of it.
- Generative Engine Optimization (KDD 2024) — the only peer-reviewed causal evidence on content modifications and generative visibility
- Google robots meta tag reference — how nosnippet, max-snippet and noindex gate AI surfaces
- Bing Webmaster Tools — first-party AI performance data, and the free ground truth for any prediction we make
- Confounding (Wikipedia) — why cited pages carrying more schema does not mean schema caused the citation
Frequently asked questions
What is the strongest single content change I can make?
Adding genuine quotations, which measured roughly +41% generative visibility in the controlled GEO study presented at KDD 2024, followed by concrete statistics at about +34%. Both only work if they are true and relevant.
Does adding schema markup get me cited more often?
There is no measured evidence that it does. A matched difference-in-differences analysis of 1,885 pages that added JSON-LD found no citation lift. Keep schema valid and consistent as hygiene, but do not budget for it as a growth lever.
Are tables really cited more often?
Tables are worth building because they chunk cleanly and answer comparison questions directly. The specific multiplier circulating in this market could not be traced to any source, so we kept the direction and deleted the number.
Does content freshness matter?
It depends on the query, and the sources genuinely conflict: one dataset shows 65% of AI bot hits going to content under a year old, another finds the median cited page is around 500 days old for evergreen questions. Audit the honest date triad, and never churn dates.
What matters most overall?
Branded mentions off your own site, which correlate roughly three times as strongly with AI visibility as backlinks do, with around 82% of citations being earned media. It is the largest factor, the slowest lever, and the one no auditor can do for you.