Most GEO advice is opinion. There is one peer-reviewed, controlled study that actually tested how on-page content edits move AI-search visibility — Aggarwal et al.'s GEO: Generative Engine Optimization (KDD 2024). It is genuinely useful and genuinely limited: a single method, an early engine generation, roughly 10,000 queries, and effects that read as correlational tendencies rather than laws. This is what it observed — and why it is best read as a starting hypothesis, not a checklist.
What the study actually did
The researchers took pages, applied nine content edits one at a time, and scored each version on a position-adjusted visibility metric against a fixed baseline. They reported that several edits raised that metric — the headline figure was up to about 40% — while one edit lowered it. That single, careful experiment is the evidence base most confident-sounding 'GEO tactics' posts are quietly built on (arXiv).
The edits it tested, and the direction each moved
Listed in plain alphabetical order — deliberately not ranked by effect — here is what the study changed on a page and the direction it reported for each:
| Content edit the study applied | Direction reported |
|---|---|
| Add attributed expert quotations | Higher visibility |
| Add statistics or quantitative data | Higher visibility |
| Adopt a more authoritative tone | Higher visibility |
| Cite credible external sources | Higher visibility |
| Improve prose fluency | Higher visibility |
| Keyword stuffing | Lower visibility |
| Make content easier to understand | Higher visibility |
| Use more unique words | Higher visibility |
| Use relevant technical terms | Higher visibility |
Eight of the nine nudged the metric up; keyword stuffing was the only edit that pushed it down — the cleanest single result in the paper. The authors reported that their larger effects tended to cluster around adding attributed quotations, statistics, citing credible sources, and improving fluency. Treat that as a loose cluster rather than a leaderboard: independent summaries of the study order those four differently, so the safe reading is 'these directions, roughly' — not 'this edit, then that one' (per-method scores, arXiv).
Does this study tell me the one change that wins AI citations?
No. It reports that several content edits — attributed quotations, statistics, citing credible sources, clearer prose — were associated with higher position-adjusted visibility in one controlled test, and that keyword stuffing was associated with lower visibility. That is directional evidence about tendencies, not a guarantee that any single edit will win you a citation.
Why to read it as early evidence
Every one of these caveats is a reason to hold the findings loosely:
- One study, one engine generation. AI search has changed repeatedly since the experiment ran. An effect measured then may be weaker, stronger, or gone now.
- A lab proxy, not your buyers. Roughly 10,000 queries against a position-adjusted metric is a controlled setup — not the live mix of ChatGPT, Gemini and Claude answering your actual prompts.
- Correlation, not proven cause. In the wild, domain authority and brand search volume are confounders, and pages that carry quotes and statistics also tend to be better pages overall. Shared features among cited pages do not prove the features caused the citation.
- Domain-dependent. The paper found the edits helped lower-authority pages more than pages already ranking at the top — the effect is not uniform across sites.
- The numbers move. AI citations swing 40–60% month to month, so any single reported effect size is a snapshot. See why point-in-time numbers lie.
Where on-page evidence stops
Even taken at face value, every edit above is about the same narrow thing — selectability, surviving the rerank once your page has already been retrieved. None of them touch discoverability, which is whether your page enters the retrieval pool at all, and which lives off-site in authority and index coverage. Polishing a page that never gets retrieved changes nothing measurable. Read discoverability vs. selectability before deciding which gate is actually failing for you, and use the Source-Gap playbook to see which one is costing you answers.
If you want a page-level read, run a free audit — it scores on-page citability signals only, and is explicit that it does not measure off-site authority (how Audit My Page works). Use the research above as a hypothesis to test against your own pages, not a recipe to follow on faith.
Sources
- GEO: Generative Engine Optimization (Aggarwal et al.) — KDD 2024 — the one controlled study. Nine content edits scored on a position-adjusted metric across ~10,000 queries; per-method scores reported in the paper.
- Zyppy Signal — AI citation ranking factors — Synthesis of 54 studies; schema ~5.6/10, llms.txt ~2.0/10. Author explicitly flags correlation, not cause.
- Search Engine Land — Chunk, Cite, Clarify, Build — Practitioner framework on self-contained passages and inline citations. Industry framework piece, not a controlled study.
- Yext original-research data, via ZipTie — Correlational industry dataset (data-rich pages earned more citation occurrences per URL). Vendor data — read as correlation, not cause.