In short
- +22 to 41%
- The range of the gain, depending on domain and topic. Quoting one number would hide the spread the study itself measures.
- What worked
- Adding statistics, quoting, naming sources. In other words: writing as though you are a source yourself.
- What worked less
- Keyword optimisation. It is not useless, but it is not the lever everyone assumed it was.
- Different per domain
- The optimal recipe differs per topic. The study found no single tactic that wins everywhere.
What does the study say?
Nine content optimisations were compared against a baseline, across several generative engines. The changes that paid off most all shared one character: they make a text repeatable. A number with a source attached can be taken over by a model without it having to assume anything.
| What was measured | How | What the study does not say |
|---|---|---|
| ~10,000 questions, several generative engines | nine content optimisations against a baseline | that this pays off equally on every topic — the gain varies per domain |
Aggarwal et al. (2024). GEO: Generative Engine Optimization. KDD '24. arxiv.org/abs/2311.09735
Why does this matter to you?
Because it points at the cheapest lever there is. You do not have to buy authority to put a sourced figure in your text — it is writing, and it is measurable. Anyone with no verifiable fact anywhere on their own site gives an engine nothing to cite with.
What does ceeme do with it?
Our fixes are built on it: a fix delivers a block of text with a figure, a source and a date, not “write more about this topic”. This site is a test of it in itself — every page here carries the weighting, the count and the provenance it talks about.
Frequently asked questions
Is this peer-reviewed?
Yes. The study was published at KDD 2024, a leading conference in this field. The preprint is on arXiv as 2311.09735.
Does 41% mean my revenue goes up?
No. It is about visibility in a generated answer, not about visitors and not about sales. Using that figure as a revenue promise means taking a correct number from a study that investigated something else.
Does this still hold, two years on?
The engines have changed, the direction has not: later studies point the same way — citable, well-supported text wins. That is also why we re-measure rather than treat this outcome as settled.
Read on
Free, no account and no card.