What is generative engine optimization (GEO)?

Who coined it, and what they actually measured

This is one of the few terms in AI marketing with a real paper behind it, so it is worth being precise about what the paper says.

Aggarwal and colleagues published “GEO: Generative Engine Optimization” in November 2023 [1]. They built a benchmark of around ten thousand queries and tested a set of content changes to see which ones increased a source’s visibility inside a generated answer. The headline finding was that several simple changes produced measurable improvement, with the best performing ones reported as lifting visibility by up to around 40 per cent on their benchmark.

The methods that worked were not exotic. Adding citations to authoritative sources. Adding statistics. Adding direct quotations. In other words, the changes that make a page more useful to quote.

The methods that did not work are the more interesting half. Keyword stuffing, the central technique of two decades of search optimisation, did not help.

Why the mechanism is different from SEO

Search optimisation competes for a position in a list. The page stays the page, and the reader clicks through to it.

Generative optimisation competes to be the material an answer is built from. The reader may never see your page. What they see is a sentence derived from it, possibly with your name attached and possibly not.

That changes what a good page looks like. A page built to rank is built to attract a click. A page built to be cited is built to be quotable: one claim per paragraph, stated plainly, with the source attached, in language that survives being lifted out of context. Those are not the same document.

The honest limits

Three, and anyone selling GEO services should volunteer all three.

The paper is one benchmark, from 2023. Models have changed repeatedly since. A technique that lifted visibility on that benchmark may behave differently now, and almost nobody re-runs the test.

Most GEO advice in circulation cites the paper without having read it. The 40 per cent figure in particular travels detached from the conditions it was measured under, which is how a benchmark result becomes a sales claim.

Retrieval does not always fire. When we put eight marketing questions to a live engine with search grounding, it only retrieved for two of them. The rest were answered from the model’s existing parameters, where no amount of page optimisation reaches. See Research.

What to do with it

1. Write one claim per paragraph and attach the source to the claim, not to the article. A paragraph that can be lifted intact is a paragraph that can be cited.
2. Put the answer in the first fifty words, before the context. That block is what gets extracted.
3. Use real numbers with dates and named sources. The paper found statistics and citations both improved visibility, and they are also just better writing.
4. Stop keyword stuffing. It did not work in the study and it makes the page worse to quote.
5. Check whether the engine is retrieving at all for your questions before investing in page changes. If it answers from memory, the work to do is different and slower.

If you run the company rather than the marketing

GEO is the most oversold term in this section, and it is oversold on the back of one real paper that most of the people quoting it have not read.

The thing worth knowing is that the techniques which worked in that study were all versions of “make the page genuinely more useful to quote”. Citations, statistics, direct quotes. The technique that failed was the one the SEO industry spent twenty years selling. That should tell you how to judge a proposal: if it sounds like gaming, it is probably the thing that already stopped working.

Ask any supplier pitching GEO two questions. What does the engine currently say about us, and did they check whether it retrieves live for our questions at all. If the answer to the second is no, page-level work will not move it, and a proposal built on page-level work is a proposal built on a wrong assumption.