A buyer asks an assistant which companies can run outbound sales for a software firm in Europe. The answer names three. There is no page two.
That is the whole problem in one sentence. Search gives you ten links and a second page nobody opens. A generative answer gives three names and nothing else.
Generative engine optimization, GEO, is the work of being one of those names. It is close to SEO and it is not the same job.
What GEO actually means
Traditional SEO gets a page ranked. GEO gets a passage quoted.
AI systems do not serve your page. They read it, pull out the part that answers the question, and write a new sentence that may or may not mention where it came from. Your unit of work stops being the page and becomes the passage.
This changes what good looks like. A page that buries its answer under four paragraphs of throat clearing ranks fine and gets quoted never. A page that answers in the first forty words gets lifted.
What the research found
The first serious study of this was published at KDD 2024 by researchers from Princeton and IIT Delhi. They built a benchmark of real queries, took the sources that answer them, and tested nine ways of rewriting those sources to see which ones made a generative engine quote them more.
Three methods moved the needle: adding statistics, citing sources, and adding quotations from named people. The paper reports improvements of roughly 30 to 40 percent on its visibility metric for the strongest methods.
One number in that paper gets repeated badly everywhere. The 40 percent is a ceiling for the best performing method on the best fitting content, not an average you should expect. Anyone selling you a flat 40 percent has not read the paper.
The finding that matters more is the direction. The things that make content quotable are the same things that make it trustworthy to a human: real numbers, named sources, and someone willing to put their name to a claim.
Source: Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, GEO: Generative Engine Optimization, KDD 2024, arXiv:2311.09735.
What this looks like on a real page
Lead with the answer. If the heading asks a question, the next sentence answers it. Everything else is support.
Put a number in it, and say where the number came from. An unsourced claim is a sentence an AI cannot safely repeat.
Write headings the way people ask things. Not Our Methodology but How long does it take to see results.
Make the page readable without running scripts. If the content only appears after four frameworks finish loading, a crawler sees an empty page.
Keep the entity data consistent. Same company name, same address, same description, everywhere you appear. An AI building a picture of who you are is matching strings.
What GEO is not
It is not a separate set of pages written for machines. Google calls that scaled content abuse and it is a good way to lose the rankings you already have.
It is not chopping your content into fragments. Google's own guidance says the opposite: use normal headings and paragraphs.
It is not a replacement for being worth citing. If three competitors have customer reviews on public sites and you have none, no amount of markup fixes that. The model is reading the whole web, not just your site.
Where to start
Pick the ten questions a buyer would actually type before choosing a company like yours. Run each one through two assistants. Write down whether you appear, who does appear, and which page they cited.
That list is your work queue. It is usually short and it is usually obvious.
We do this as a service for IT and SaaS companies, and the work is set out on our SEO, GEO and AI visibility page.


