What Is Generative Engine Optimization (GEO)? What the Research Found
Written by Armen Andonian · Published on 28 September 2026
Here is a number that should change how you write a web page. When researchers at Princeton and IIT Delhi added a few relevant quotations to ordinary web pages, those pages took up about 41 percent more of the AI generated answers built from them. Stuffing the same pages with keywords made them less visible, not more.
That study is where the phrase generative engine optimization comes from. I am Armen, an SEO consultant in Barcelona, and I run the agency ACERO Digital with a team of specialists. Most of what gets sold as GEO today quotes the headline figure and skips the paper, so this post goes back to what was actually tested, including the parts nobody should be selling you as proven.
What is generative engine optimization?
Generative engine optimization (GEO) is the practice of shaping your content so that an AI system writing an answer uses more of it and credits you as the source. The target is a larger, earlier share of the generated response with your link attached, not a rank in a list.
The key word is generative. A featured snippet lifts your sentence as it stands. ChatGPT, Perplexity, Gemini and Google’s AI Overview read several pages and write something new, so your text gets rewritten and sometimes dropped entirely.
GEO is about surviving that blender with your name still on it.
Where does the term GEO come from?
It comes from a single paper: “GEO: Generative Engine Optimization” by Pranjal Aggarwal, Vishvak Murahari and colleagues, posted in November 2023 and published at KDD 2024, one of the main data science conferences. Before it, people talked loosely about “optimising for AI”. After it, there was a name and a way to measure the thing.
The team built GEO bench, a set of 10,000 queries covering domains as different as law and history. For each query, a test engine pulled the top Google results, then wrote a cited answer from them. The researchers rewrote one of those source pages at a time and watched how much more, or less, of the answer it earned.
That setup matters. Every number below comes from changing the content of a page that was already retrieved. None of it measures how you get retrieved in the first place.
How do you measure visibility in an AI answer?
The paper used two measures, and both are more useful than the “AI visibility score” on most dashboards.
The first is position adjusted word count: how many words of the answer are attributed to your source, weighted so that a citation near the top counts more than one buried at the end. The second is subjective impression, where a model rates your source’s presence on things like relevance, influence and how likely a reader would be to click it.
You can run a rough version of the first measure by hand. Ask a question in Perplexity, count the sentences that cite you, note where they sit. It is crude, and it is still more honest than a single score to two decimal places.
Which GEO methods actually worked?
Adding quotations, statistics and cited sources worked best; keyword stuffing was the only method that lost ground. Here are the nine methods the paper tested, with the relative change against an unedited page.
| Method | Position adjusted word count | Subjective impression |
|---|---|---|
| Quotation addition | +41% | +28% |
| Statistics addition | +33% | +20% |
| Fluency optimisation | +29% | +13% |
| Cite sources | +28% | +13% |
| Technical terms | +19% | +9% |
| Easy to understand | +14% | +5% |
| Authoritative tone | +12% | +16% |
| Unique words | +6% | +3% |
| Keyword stuffing | 8% lower | +5% |
Read that table as a writer, not a marketer. The winners all add evidence a reader can check: a named person saying something, a figure with a source, a reference to where a claim came from. The losers add decoration.
When the team repeated the tests on the live Perplexity engine, the pattern held, though the gains were smaller. Quotation addition lifted the main metric by about 22 percent there.
Why do lower ranked pages gain the most from GEO?
Because the engine already leans on the top result, so a lower ranked page has far more room to grow. In the paper, citing sources lifted a page ranked fifth on Google by 115 percent, while the same change on the first ranked page cut its visibility by about 30 percent.
This is the finding I find most useful for small businesses. If a big directory or a national chain holds position one for your service in Barcelona, you will not outrank it this quarter. You can still become the page the answer quotes, because the engine cares about which text is easiest to cite with confidence, not only about who ranks highest.
It cuts the other way too. If you already rank first, piling on citations is not a free win.
Does GEO work the same in every industry?
No, and the paper says so plainly. Statistics helped most on law and government questions. Quotations and an authoritative tone did well on history. Plain source citations helped most on straightforward factual queries.
For a business that means testing your own queries rather than copying a checklist. A dental clinic in Eixample answering “how much does an implant cost in Barcelona” needs a real price range with a date on it. A law firm explaining the Beckham regime needs the article of the law it rests on. Different evidence, same principle.
What does the research not prove?
It does not prove that GEO gets you retrieved, and that gap is where a lot of paid GEO work goes wrong. Every test started from pages the engine had already selected. If your site is blocked to crawlers, thinly indexed, or simply not among the pages an engine pulls for the query, none of these edits get a chance to work.
It was also run on a test engine, then checked on one live product, in 2023 and 2024. Models and retrieval systems have changed several times since. The direction of the findings (evidence beats decoration) has aged well; the exact percentages should be treated as a guide, not a promise.
That is why a real GEO engagement starts with ordinary groundwork. A technical SEO audit confirms that search bots and AI crawlers such as OAI-SearchBot and PerplexityBot can reach your key pages. Only then is rewriting worth the money.
How do you apply GEO to your own site?
Start with the pages that should answer your customers’ buying questions, and add checkable evidence to them.
- List fifteen or twenty questions a customer asks before hiring you, and run them in ChatGPT and Perplexity, then in Google. Note which pages get cited.
- Pick the one page of yours that deserves to be cited and is not.
- Add a real statistic with its source and year in the same sentence, and a short quotation from a named person, whether a customer, a regulator or one of your own specialists.
- Put the direct answer in the first two sentences under each heading, in plain language.
- Remove any sentence that repeats the keyword without adding a fact.
- Recheck the same questions after the page has been recrawled, and log the results monthly.
This is most of what we do in an SEO consulting engagement focused on AI search. It is editorial work more than technical wizardry, and it tends to make the page better for human readers at the same time.
Where I would start
If I had one hour with your website, I would not open a GEO tool. I would pick your most important service page and ask a blunt question: is there a single fact on it that a stranger could verify? No price, no date, no name, no source usually means no citation.
Fix that first. Add a real number and a real quote, each linked to where it came from, then watch what Perplexity does with the page a few weeks later. If you want us to look at which of your pages AI answers are skipping and why, send me the questions your customers ask and we will start there.
Frequently asked questions
What is generative engine optimization in simple terms?
Generative engine optimization, or GEO, is the work of making your content more likely to be used and credited when an AI system such as ChatGPT, Perplexity or Google's AI Overview writes an answer. The goal is a visible share of the generated answer with your source attached, rather than a position in a list of links.
Who coined the term generative engine optimization?
The term comes from a research paper by Pranjal Aggarwal and colleagues from Princeton University and IIT Delhi, first posted in November 2023 and published at the KDD 2024 conference. It introduced the name, a benchmark of 10,000 queries called GEO bench, and the first measured tests of which content changes raise visibility inside AI generated answers.
Does keyword stuffing work for GEO?
No. In the original GEO study, keyword stuffing lowered visibility by about 8 percent on the main metric compared with leaving the page alone. Adding quotations, statistics and cited sources did the opposite, with gains of roughly 28 to 41 percent on the same measure.
Is GEO different from SEO?
GEO builds on SEO rather than replacing it. Most AI answers are assembled from pages a retrieval system has already found and ranked, so crawlability, indexing and authority still decide whether you are in the pool. GEO changes what happens after that point: how easy your text is to quote and whether the engine keeps your name attached when it rewrites it.
How long does generative engine optimization take to show results?
For engines that search the web live, such as Perplexity or ChatGPT with search, a rewritten page can be picked up as soon as it is recrawled, often within weeks. What a model remembers from its training data moves far slower, on the timescale of new model releases, so treat that part as a long game.
Armen Andonian
SEO Consultant in Barcelona
I'm an SEO consultant in Barcelona and founder of ACERO Digital. I help local businesses, ecommerce and international companies rank on Google and win customers from search.
LinkedIn →