Generative engine optimisation, or GEO, is the work of making a web page more likely to be used and credited in an answer that an AI tool writes, such as ChatGPT, Perplexity or Google’s AI Overviews.
The name did not start as a sales term. It is the title of a research paper first posted in November 2023 by six researchers, four of them at Princeton University and the Indian Institute of Technology Delhi, and accepted to KDD 2024, a computer science conference. The paper says exactly what was tested, and that is narrower than the advice now sold under its name.
Where Does the Name Come From?
The paper is called GEO: Generative Engine Optimization, by Pranjal Aggarwal and five co-authors. Its first version appeared in November 2023 and its third in June 2024. By a generative engine the authors mean a search tool that answers by writing from several sources instead of listing links. GEO is their name for the changes a website owner can make to be seen more in those answers. Agencies now sell the same idea as GEO SEO.
To test the idea they gathered 10,000 questions from nine sources, from real search logs to questions written by another AI. One set is essay questions from All Souls College, Oxford, which is the only place our city appears in the paper.
What Did the Researchers Actually Test?
They built an answer engine of their own. For each question it took the top five results from Google and had an OpenAI model, GPT-3.5, write an answer from them. They then picked one of the five pages at random, had a language model rewrite it in nine different ways, one at a time, and measured how much of the answer came from that page.
Their first measure counted the words in the answer that were credited to the page, giving more weight to sentences near the top, where a reader is most likely to look. They checked the same changes on Perplexity, an answer engine people actually use.
- Making the tone more authoritative and persuasive.
- Replacing general statements with statistics.
- Stuffing in more of the words from the question, the old SEO habit.
- Citing sources.
- Adding quotations from credible sources.
- Making the language easier to understand.
- Making the writing more fluent.
- Adding unusual words, and adding technical terms.
What Did It Find?
Three changes did best: citing sources, adding quotations and adding statistics. On the word-count measure they made a page 30 to 40 per cent more visible in the test engine’s answers. Plainer, more fluent writing helped too, by 15 to 30 per cent.
Two things people often try did not help. A more authoritative tone made no significant difference, and stuffing in keywords did little or nothing. On Perplexity the pattern held: adding quotations did best, at 22 per cent, and keyword stuffing left a page 10 per cent worse off than doing nothing.
| Change to the page | What happened to its visibility |
|---|---|
| Citing sources, adding quotations or adding statistics | Up 30 to 40 per cent |
| Easier or more fluent writing | Up 15 to 30 per cent |
| A more authoritative tone | No significant improvement |
| Keyword stuffing | Little or no improvement |
Who Did It Help Most?
The most useful result is in one of the paper’s smaller tables. It splits the gains by where the rewritten page sat in the search results, and the authors sum it up in one line: “GEO is especially helpful for lower ranked websites.”
For a page in fifth place, citing sources more than doubled its share of the answer, an increase of 115.1 per cent. For the page in first place, the same change cut its share by 30.3 per cent. Adding quotations and adding statistics showed the same shape. A small business is more often fifth than first, so in our judgement that is the encouraging half of the finding.
What Does the Paper Not Show?
It is a careful study, and its limits are part of reading it well. These are the ones we would keep in mind before paying for anything sold as GEO:
- Its engine was a 2023 model reading five Google results, not ChatGPT or Google’s AI Overviews as they work today.
- Most of its questions looked for information; one in ten was the kind somebody asks when ready to buy, such as a search for an agency in a town.
- Each page was rewritten alone while the other four stayed as they were. In a real market, competitors change their pages too.
- It measured how much of the answer came from a page, not visits, enquiries or sales.
- The authors found the effect varied by subject, so a change that worked for history questions may not work for yours.
How Does It Compare With What We Saw in Oxford?
One detail of the method matters more than any of the percentages. The test engine only ever read pages that were already in Google’s top five for the question. A page outside them was never in the answer, however well it was written.
That matches what we found when we read Google’s AI answers for Oxford searches in August 2026: most of the agencies they named already had a website ranking on page one of the ordinary results. As it was tested, GEO starts where ordinary SEO ends. It is a way to be quoted more once you are in the room, not a way into the room.
What Should You Do With It?
Treat GEO as a set of writing habits added to ordinary SEO, not as a separate service. In our judgement, these are the habits the paper supports:
- Earn a place near the top of the ordinary results first; the test engine never looked past the fifth.
- Back up what you claim: say where a figure comes from, and link to the source when you can.
- Quote people a reader could check, such as a named client or a published expert.
- Use numbers you can stand behind in place of words like leading and trusted.
- Write plainly, and leave keyword stuffing alone; it did nothing, or worse.