Does generative engine optimisation work? What the research found

The term comes from a 2023 research paper

Generative engine optimisation (GEO) is the work of making a page more likely to be used in an answer written by an AI engine. The term comes from a paper called "GEO: Generative Engine Optimization" by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, from Princeton University and IIT Delhi (arXiv, 2023). It was first posted on 16 November 2023 and accepted to KDD 2024.

A generative engine is a search tool that writes one answer from several web pages, such as Perplexity or Google AI Overviews. The authors built GEO-bench, a benchmark of varied search queries across many subjects. A benchmark is a fixed set of tests that lets researchers compare methods on equal terms. They then rewrote pages in nine ways and measured how much each change moved the page's share of the answer.

The nine methods were:

• Authoritative: a more confident, expert tone
• Statistics Addition: adding figures
• Keyword Stuffing: repeating the search terms
• Cite Sources: adding references to other sources
• Quotation Addition: adding quotes from people
• Easy-to-Understand: simpler wording
• Fluency Optimization: smoother sentences
• Unique Words: less common words
• Technical Terms: more specialist vocabulary

They measured how much of each answer drew on a page

The main measure counted how much of the AI answer came from a given source. The paper calls it Position-Adjusted Word Count. Words near the top of the answer count for more, because readers see them first. A page that supplies the opening lines scores higher than one that supplies a line at the end.

The second measure was a subjective impression score. It rated how strongly a source came through in the answer, as a reader would experience it. The two measures did not always move by the same amount, so the paper reports both.

Quotes, statistics and citations helped most

Three methods did best: Quotation Addition, Statistics Addition and Cite Sources. They raised visibility by roughly 30% to 40% on the word-count measure and 15% to 30% on the impression measure. The paper's headline figure is "up to 40%". On Perplexity, Quotation Addition raised visibility by 22%.

The three edits share one trait. Each adds something an engine can lift into an answer and attribute: a figure, a named person's words or a source. Here is how each could look on a local service page. The business and every detail in the table are made up.

None of these edits is a trick. Each makes the page more specific and easier to check. We write service pages this way for that reason alone.

Keyword stuffing made pages less visible

Repeating the search terms did worse than leaving the page alone. Keyword stuffing is the old habit of packing a phrase such as "plumber Parramatta" into a page as often as possible. In the study it scored below the unchanged page. On Perplexity it made the page 10% less visible.

So if a page reads as though it was written for a search engine, the research gives no reason to keep it that way. Write the phrase once where it belongs, then spend the words on facts.

Lower-ranked pages gained the most

The edits helped pages further down the search results more than the pages at the top. Cite Sources raised the visibility of the 5th-ranked result by 115.1%. Over the same tests, the top-ranked result's visibility fell by 30.3%.

For a small business, this is the most useful finding. Large sites often hold the top organic spots. The study suggests a well-sourced page further down can still take a bigger share of the AI answer. It does not show that this happens on every query, or that it lasts.

The study can't tell you how today's engines behave

The paper is careful evidence, and it has clear limits. Keep these in mind before you rely on any one figure:

• It was a lab benchmark. The pages and queries were a fixed test set, not live search traffic.
• The main engine was built on GPT-3.5 and read the top 5 Google results. Perplexity was the only live product tested.
• Engines have changed a great deal since 2023. Google AI Overviews, AI Mode and ChatGPT search work in ways the paper did not test.
• The edits were applied to one page at a time. When every page in a result adds statistics, the gain for any one page may shrink.
• GEO-bench covers many kinds of question. It was not built around local service searches, so we can't say how well the figures carry over to them.

We treat the results as a sign of which way to lean. We do not treat them as a forecast for a given business.

What we take from it in practice

We write pages that an engine can quote with confidence. That means four habits, and none of them hurts a human reader:

1. Answer the question in the first two sentences. Our guide to answer-first pages shows how.
2. Use real figures, and say where each one comes from and when.
3. Quote people with standing, such as a licensed tradesperson, a named expert or an official body.
4. Keep business facts the same everywhere: name, address, services, licence and hours.

The paper's methods are only one part of the picture. Engines must be able to reach and index a page before they can use it. Google says a page must be indexed and eligible to show with a snippet to appear in AI Overviews or AI Mode (Google Search Central). Our guide on how generative engine optimisation works covers that side.

For a live example of AI citations, look at Gridly, a separate Australian guide to EVs and home energy. Liam built its SEO and content system with the Gridly team. In October 2026, Ahrefs showed 1.7K AI responses across platforms and 508 AI Overview citations for the site. We can't say which parts of the work caused those citations, and we don't credit them to the paper's methods. The Gridly case study has the details.

If you want this applied across a whole site, our generative engine optimisation service starts with the pages your customers read first.