Generative engine optimisation: the paper that named it, and what it actually tested
GEO comes from a 2023 arXiv paper, not a standards body. What the paper measured, and a prompt that separates the methods it tested from vendor invention.
- Works in
- ChatGPT, Claude, Gemini
- You need
- The full text of one page, headings included · The method list from a vendor proposal, if you have one
- Written for
- what is generative engine optimization geo
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
Generative engine optimisation is the one of the two labels with a document behind it. The term was introduced in a named research paper, not by a standards body and not by any AI operator, and the difference matters because a paper can be read, checked and disagreed with. The prompt below holds your page and your vendor proposal against what that paper actually tested.
What the paper is, and what it measured
GEO: Generative Engine Optimization was submitted to arXiv on 16 November 2023 by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, and accepted to KDD 2024.
Its framing is the useful part. It names website owners as the third stakeholder in generative search, alongside the engine and the user, and argues they have almost no control over when and how their content appears. That is the problem the whole category is a response to.
It then tested nine content methods and reported that its best performing one raised visibility by up to forty per cent. Two qualifications travel with that number and are usually dropped. It was measured against a simulated generative engine, built on GPT-3.5-turbo reading the top five Google results, and scored on a benchmark the authors built themselves. The paper also states that effectiveness varied by domain.
None of that makes the result weak. It makes it a result about a specific setup rather than a promise about ChatGPT.
What the nine methods tell you
The tested list is short enough to hold in your head: Authoritative, Keyword Stuffing, Statistics Addition, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words, Technical Terms.
Read it once and most GEO proposals become easier to price. Two of the nine, citing sources and adding real quotations, are simply good editing and worth doing on grounds that have nothing to do with any engine. One of them is keyword stuffing, included as a tested condition rather than a recommendation.
Everything a vendor sells you outside those nine was not tested by the paper. It may still be sensible. It is not evidence, and the prompt refuses to let the two blur.
Why the prompt sorts a proposal
Ask a model what GEO is and you get a tidy definition with a standards body quietly invented into it. This one is pinned to the paper, forced to attach the evaluation setup to the headline figure, and made to sort every line of a vendor scope into tested, plausible or unsupported.
For the other half of the argument, see answer engine optimisation defined on its own terms and AEO and GEO, which is which. For the retrieval side the paper does not cover, see how LLM SEO actually works.
You are a research literate content strategist. Generative engine optimisation
is a term introduced in a specific academic paper. I want you to hold my page,
and any vendor method list I paste, against what that paper actually did.
Ground rules you must follow exactly.
1. GEO was introduced in "GEO: Generative Engine Optimization" by Aggarwal,
Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, posted to arXiv in
November 2023 and accepted to KDD 2024. It is a research contribution. It
is not a specification, a standard, or guidance published by any AI
operator. Say this once and do not soften it.
2. The paper evaluated nine named content methods: Authoritative, Keyword
Stuffing, Statistics Addition, Cite Sources, Quotation Addition,
Easy-to-Understand, Fluency Optimization, Unique Words, Technical Terms.
Treat that list as the boundary of what was tested. Anything outside it was
not tested by the paper, whoever is selling it.
3. The headline figure was measured against a simulated generative engine
built on GPT-3.5-turbo reading the top five Google results, on the authors
own benchmark. Whenever you refer to it, attach that setup. Never present
it as a result observed in a live consumer assistant.
Now do three things.
PART ONE, my page against the tested methods. Go through the nine methods in
order. For each one, say whether my page already does it, quote the exact
passage that shows it or write "absent", and give one concrete change. For
Keyword Stuffing, say that it was included as a tested condition and that
copying it is a bad idea on every other ground, and move on.
PART TWO, the vendor list. If I pasted one, sort every item into exactly one
of three buckets and name the bucket for each:
- Tested by the paper. Name which of the nine it corresponds to.
- Plausible but untested. Reasonable, no evidence in the paper either way.
- Unsupported. Contradicted, unfalsifiable, or a claim about mechanics no
operator documents.
Do not invent a fourth bucket to be kind. If I pasted nothing, skip this part.
PART THREE, the limits, in plain language. Three to five sentences on what the
paper cannot tell me: that it studied a simulated engine rather than the
assistant my buyers use, that method effects varied by domain in the paper
itself, and that no operator has published how a live assistant selects
sources.
Constraints:
- Do not invent statistics, percentages or adoption figures.
- Do not attribute GEO to a standards body, an operator or a specification.
- Do not present GEO and AEO as a settled taxonomy. Usage is inconsistent.
- If my page is too short to judge against a method, say so rather than
inventing a passage.
My page: [PASTE THE FULL PAGE TEXT INCLUDING EVERY HEADING]
My vendor method list: [PASTE THE DELIVERABLES OR METHODS FROM A PROPOSAL, OR WRITE "none"]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[PASTE THE FULL PAGE TEXT INCLUDING EVERY HEADING]- One real page with its headings intact, pasted rather than linked. Choose a page you would defend to a buyer, not the homepage. The nine methods are judged against passages, so a page pasted as unbroken text gets judged as one long passage and every verdict comes back vaguer than it needed to be.
[PASTE THE DELIVERABLES OR METHODS FROM A PROPOSAL, OR WRITE "none"]- The actual bullet list from an agency deck, a tool feature page or a scope of work. This is the input that makes part two worth running, and it is the fastest way to find out how much of what you are buying was ever tested by anybody. Writing "none" skips that section cleanly.
How to run it
- 01Read the abstract yourself before you run anything
Open arxiv.org/abs/2311.09735 and read the abstract. It takes two minutes and it inoculates you against the version of GEO you will be sold, which is usually the word without the study. You are about to ask a model about a document, so know what the document says.
- 02Get the vendor method list in writing
Copy the deliverables from the proposal, deck or product page verbatim, including the vague ones. Do not tidy them up. Part two sorts them into tested, plausible and unsupported, and a tidied list hides exactly the items that would land in the third bucket.
- 03Paste one page, headings included
Copy the visible text of a single page with the heading structure intact. Do not hand the model a URL. Some assistants cannot fetch it, others fetch a rendered version that differs from what a crawler receives, and the fetched copy usually arrives with the headings flattened.
- 04Check the paper facts in the output before you read the advice
Confirm it did not promote the paper to a standard, did not attribute GEO to Google or OpenAI, and did not quote the headline figure without the simulated setup attached. Models are strongly inclined to tidy a research result into a rule. If it did any of those, distrust the rest of the output.
- 05Act on Cite Sources and Quotation Addition first
Of the nine, those two are the ones that are also plainly good editing: linking the primary source you actually used, and quoting a named person or document rather than paraphrasing. They improve the page for a reader whether or not the paper generalises to a live assistant, which makes them the safest place to spend an afternoon.
- 06Measure the page rather than trusting the method list
Run the edited page through the free AI content readiness check. It scores whether sections are bounded by headings, whether answers arrive in the first line and whether an author is named. A page that ticks nine methods and still scores badly has been decorated rather than restructured.
Questions people ask
What is generative engine optimisation?
It is the practice of changing a website so it is more likely to be used and cited when a generative system writes an answer. The term was introduced in a 2023 research paper that proposed a framework for measuring visibility inside a generated response. It is a research paradigm and an industry label, not a standard anybody is obliged to follow.
Where does the term GEO come from?
From "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, submitted to arXiv on 16 November 2023 and accepted to KDD 2024. The paper introduced both the term and GEO-bench, a benchmark of user queries used to evaluate it.
What did the GEO paper actually test?
Nine content level methods: Authoritative, Keyword Stuffing, Statistics Addition, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words and Technical Terms. Effectiveness varied by domain, which the paper itself states. Anything a vendor sells you outside those nine was not tested by the paper, whatever it is called.
Does the 40% improvement figure apply to ChatGPT?
It was not measured on ChatGPT or on any live consumer assistant. The paper evaluated against a simulated generative engine built on GPT-3.5-turbo reading the top five Google results for each query, scored on visibility metrics and a benchmark the authors built themselves. It is a real result about that setup and it is not a promise about a product.
Is GEO different from AEO?
Only one of the two has a document behind it, and beyond that there is no agreed boundary. Some practitioners treat AEO as the older direct answer surfaces and GEO as the generated ones. Others use one as the umbrella for the other. Neither reading is wrong, because no authority exists to settle it.
Unsubscribe in one click. We never pass your address on.