Generative engine optimization (GEO), explained

Who do AI engines cite when you ask what GEO is?

The fastest way to understand generative engine optimization is to watch it happen on the question itself. Attensira tracks "what is generative engine optimization" as one of the prompts in its own workspace and reads the answers on ChatGPT and Perplexity on a daily schedule. In the 30 days from 4 September to 3 October 2026, read over the Attensira MCP on 4 October, that produced 102 answers, 57 on ChatGPT and 45 on Perplexity. The table at the top of this page is who those answers leaned on, counted as answers that cited each source.

Three things are worth noticing, because they are GEO in miniature.

The engines disagree about where the answer lives. Wikipedia is the one source both engines rely on. Beyond it, ChatGPT and Perplexity barely share a top source for the same question on the same days. A page that wins one engine can be invisible on the other.

Academic sources carry weight on one engine and barely register on the other. ChatGPT cited princeton.edu in 33 of its 57 answers and arxiv.org, where the paper that coined the term is listed, in 20. Perplexity cited them in 2 and 4 of 45. Which kind of source an engine trusts for a definition is itself something to measure, not assume.

Being named and being cited are different events. Semrush was named in 22 of the 102 answers, all of them on Perplexity, while its own pages were not among the most cited for this question on ChatGPT. Attensira, the company publishing this page, was named in none of the 102 answers. That is our baseline on our own category's definition, and it is the reason this page exists.

Where does the term generative engine optimization come from?

From a paper, which makes it unusual in this field. GEO: Generative Engine Optimization is by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. It was first submitted to arXiv on 16 November 2023, revised in May and June 2024, and published at KDD 2024.

The neighbouring terms, answer engine optimization, AI SEO, LLM SEO and LLMO, came from practitioners and vendors. Wikipedia's article on the subject notes that these overlapping terms are used for related practices and that usage varies with no consensus definition. When you need a definition you can point at, GEO is the one with a citation behind it.

The problem the paper names is a change of audience. In classic search, a publisher competes for a place in a list that a person reads. In a generative engine, retrieved pages are summarised into prose, and a page either survives into that prose, attributed and near the top, or it does not appear at all. The authors set out to define what visibility means inside a generated answer, to build a benchmark large enough to measure it, and to test whether specific edits move it.

A source citation is the link an engine attaches to an answer, and the event GEO measures. Grounding is how an engine ties its answer to retrieved pages. Semantic search is the retrieval method that finds those pages by meaning rather than exact words. AI search is the wider shift from ranked links to written answers.

What did the GEO paper measure?

A benchmark. GEO-bench holds 10,000 queries, split into 8,000 for training, 1,000 for validation and 1,000 for testing. They come from nine datasets and span 25 domains, from arts and health to games, law and history.

Two visibility metrics. The first, Position-Adjusted Word Count, counts the words of the sentences attributed to a source and discounts them the further down the answer the citation sits. A source quoted at length near the top scores high; a source mentioned once at the bottom scores low. The second, Subjective Impression, asks a model to rate a source's presence on seven aspects: relevance, influence, uniqueness, subjective position, subjective count, likelihood of a click and diversity.

Nine edits. Each method is a way of rewriting a source page: making it more authoritative in tone, adding statistics, stuffing keywords, citing sources, adding quotations, making it easier to understand, improving fluency, adding unique words, and adding technical terms. The researchers rewrote pages each way and measured how the generated answers changed.

Most of the experiments ran against a generative engine the researchers built for the purpose. A subset ran against Perplexity, a commercial engine, which is the closest the paper comes to the systems you care about.

Which GEO methods worked, and which did not?

On Position-Adjusted Word Count, the three strongest edits were:

• Quotation Addition

  What it means on a page

  Quote a named, relevant source in the text

  Score on the paper's main metric

  27.8

• Statistics Addition

  What it means on a page

  Replace vague claims with specific numbers

  Score on the paper's main metric

  25.9

• Cite Sources

  What it means on a page

  Attribute claims to their sources in the text

  Score on the paper's main metric

  24.9

• Keyword Stuffing

  What it means on a page

  Repeat the query's terms

  Score on the paper's main metric

  17.8, below the unedited page

The abstract summarises the headline result as GEO boosting visibility by up to 40% in generative engine responses. On Perplexity, adding statistics improved the Subjective Impression metric by 37%.

The keyword stuffing row is the most useful line in the paper. A decade of search habit says that repeating the query's words helps. In generated answers it did worse than leaving the page alone. What helped was evidence: numbers, quotations and named sources. That is a change in what good writing for search means, and it agrees with what most careful editors would have said anyway.

The best method depends on the subject. The paper is explicit that efficacy varies across domains. An authoritative tone helped most on debate and history questions. Citing sources helped most on statements of fact. Quotations helped on people and society and on history. Statistics helped on law and government and on opinion questions. There is no single GEO checklist that the evidence supports for every page.

Who benefits most from GEO?

The paper's most surprising finding is about who gains. Citing sources raised visibility by 115.1% for sites ranked fifth in the underlying search results, while on average the top-ranked site lost 30.3% of its visibility. The authors read this as GEO helping smaller publishers compete with the sites that already dominate search.

Treat that as a promising result from one benchmark, not as a law. It suggests that in generated answers, evidence on the page can outweigh rank in the list the engine retrieved. It does not tell you that a fifth-ranked page in your market will overtake the leader.

What does the GEO paper not prove?

A careful reading matters here, because most writing about GEO overstates it.

It does not say any commercial engine uses these rules. The experiments ran mainly on a constructed engine, with a subset on Perplexity. Nothing in the paper claims that ChatGPT, Gemini or Google's AI features reward these edits, and none of those platforms has said so. Google's own documentation says there are no additional requirements and no special optimizations needed to appear in AI Overviews or AI Mode, beyond a page being indexed and eligible for a snippet.

It measures presence in an answer, not business results. Visibility in the paper is words and position inside a generated response. It does not measure whether anyone clicked, signed up or bought.

It is one benchmark at one point in time. Engines change their models and retrieval often. Effect sizes measured in 2023 and 2024 are a direction to test, not a constant.

It does not cover what a model already remembers. The paper's engine wrote answers from pages it retrieved for each query. Commercial assistants also answer from what the underlying model learned in training, sometimes without retrieving anything. A brand a model already knows can be named with no source attached, and no edit to a page changes that in the short run. GEO works on the retrieved part of the answer.

It does not cover off-site sources. The paper edits the source pages themselves. In practice, as the table at the top of this page shows, most of what an engine cites belongs to someone else: Wikipedia, course pages, review sites. GEO in the paper's sense is about your own pages. Winning an answer usually also needs work on other people's.

How do you apply GEO to a page?

Start with the pages the engines already retrieve. These are ours, the attensira.com pages cited most in the answers to our 61 questions over the 30 days to 4 October 2026, read over the Attensira MCP that afternoon:

Six of the seven were cited only for questions that already contain our name, and the two homepage rows are one page counted under two hosts. The pricing page is the exception: an engine retrieves it for an unbranded question and then answers without our name. That is where a GEO edit pays first, on a page the engine already reads for a question a buyer asks.

Then take the question the page should answer and rewrite the passage that answers it, following the evidence.

1. Answer first. Put a plain, complete answer to the question in the first two or three sentences, so the passage still makes sense when an engine lifts it on its own.
2. Replace adjectives with numbers. "Fast setup" becomes the number of minutes, measured, with the date. "Trusted by many" becomes nothing at all until it can be a sourced number.
3. Name your sources in the sentence. "According to Google's Search Central documentation" carries more weight than a bare claim, and it is the edit the paper found worked.
4. Quote, where a quote is the evidence. A short quotation from a named, relevant source is the single strongest edit in the paper.
5. Remove the keyword repetition. Write the page for the person asking. Repeating the query did worse than doing nothing.
6. Check the page can be fetched and quoted. No edit matters if the engine's crawler is blocked or the passage is hidden behind a script. The AI crawler guide covers access.

Here is the shape of one such edit, on a passage about pricing.

Before: "Our flexible plans fit teams of every size, with powerful features and great value."

After: "Attensira's plans differ by how many prompts they track and how many credits they include each month. Agent work is paid in credits, and more credits can be bought at any time, so a quiet month costs less than a busy one. The pricing page lists each plan's prompts, included credits and the price of a top-up, dated."

The second version answers a question, states how the pricing works, and points at the source. It is the passage an engine can quote without inventing anything.

Why do ChatGPT and Perplexity cite different sources for the same question?

Because they are different products built on different retrieval. Each engine runs its own search over its own index, or over a partner's, and each has its own habits about how many sources to attach and which kinds it prefers. The table at the top of this page is one question, read on the same days, and the two engines' favourite sources barely overlap. That is the normal case, not an exception.

For GEO this matters in two ways. First, an edit to your own page only helps on the engines that retrieve it. If Perplexity answers a question from Wikipedia, Coursera and Built In, a better passage on your site may never be read there, however well it follows the paper. Second, the edits the paper measured were tested mostly against one constructed engine. How much each edit helps on each commercial engine is an open question, and the only way to answer it for your own pages is to measure each engine separately.

The practical rule is to start from the answers, not from the page. Read which sources each engine leans on for the questions you care about. Where your own page is among them, GEO edits to that page are the lever. Where it is not, the first job is to be retrievable at all: crawlable by that engine's search agent, indexed, and answering the question plainly enough to be selected. Where the deciding sources are other people's pages, the lever is getting onto those pages, which is outside what the paper studied.

What should a GEO review of one page check?

A review that follows the evidence asks a short list of questions about the page, in this order.

1. Which questions is this page meant to answer? Write them down in the words a buyer would use. A page that is meant to answer everything is selected for nothing.
2. Can each engine's search agent fetch it? Check robots rules for the search crawlers of the engines you care about, and check that the answer is in the server-rendered HTML rather than injected by a script.
3. Does the answer come first? The opening two or three sentences under each heading should answer the heading's question without the rest of the page.
4. Are the claims evidence or adjectives? Every claim that could carry a number should carry one, with its source and date. Every superlative without a number should go.
5. Are sources named in the sentence? A claim that names where it comes from is the "cite sources" edit. A footnote nobody reads is not.
6. Is there a quotation where a quotation is the best evidence? A short quote from a named expert, a standard or a primary source.
7. Is anything repeated for the sake of the query? Remove it. The paper's keyword stuffing result is the reason.
8. Which other pages decide this answer? If the answer leans on third-party pages, list them. That list is the off-site half of the work.

None of these needs special markup or a new file. They are editorial checks, which is what the evidence says GEO is.

Does GEO apply to every kind of question?

Not equally. The paper's own domain results suggest the right edit depends on the kind of question. Factual questions rewarded citations. Questions about people and history rewarded quotations. Opinion and policy questions rewarded statistics. Debate and history rewarded an authoritative tone.

Commercial questions, the kind a buyer asks before choosing a product, were not the paper's focus, and they behave differently again. Answers to "best tool for X" questions tend to lean on comparison articles, listicles and review sites, and they name several products. On those questions, the evidence on your own page competes with what other people have written about you, and the off-site work usually matters more than any edit to your own site. In our own workspace, the answers to the 61 questions we track carried 97,149 citations in the 30 days to 4 October 2026, and 1,442 of them pointed at attensira.com. The other citations went to review sites, documentation, competitors and publishers. That is why Attensira's crew spends as much time on the pages that decide an answer as on our own site, and why a GEO plan that stops at your own pages leaves most of the answer untouched.

How is GEO different from SEO and AEO?

• The goal

  Classic SEO

  A high position in a ranked list

  GEO

  Inclusion and prominence inside a generated answer

  AEO

  Being the passage an answer quotes

• The unit

  Classic SEO

  A ranking for a keyword

  GEO

  Words and position attributed to your source

  AEO

  One extractable answer

• The evidence base

  Classic SEO

  Two decades of practice and search engine guidance

  GEO

  One peer-reviewed paper and its benchmark

  AEO

  Practitioner experience

• Where the work happens

  Classic SEO

  Your site, plus links to it

  GEO

  Mostly your own pages, per the paper

  AEO

  Your own passages

The honest summary is that they overlap heavily. GEO's contribution is evidence: a measured answer to which edits change visibility in a generated answer. The broader practice of getting named and cited across every engine, on and off your own site, is what most people now call AI SEO. Answer engine optimization is the older, passage-level view of the same work.

How do you know whether GEO worked?

Not by asking an engine once. Generated answers are not deterministic: the same question asked twice can return different sources and different names.

Measure it the way the paper did, at a smaller scale. Pick the questions the edited page should answer. Ask each one on each engine every day for a couple of weeks before the change and a few weeks after. Count, per engine, how often your brand is named and how often your domain is cited, and keep the sample size beside every number. Call it a change only when the difference is larger than the noise.

Attensira does this as a matter of course. After a change ships, the answers are re-read at 7, 14 and 30 days, and each re-read gets a verdict with its sample size, where "no change" and "too close to call" are verdicts too. The GEO measurement guide explains how to build the question set and read the numbers yourself.

What does Attensira do with GEO?

Attensira is an AI-native SEO agency run by software: the service of an SEO agency at the price of a tool. On GEO, the crew works the way the paper measured: one edit to one source, judged by the answers before and after. When a tracked question is lost, the Researcher finds the passages and pages the engines lean on, and the Copywriter and Writer draft the edits the paper found worked, such as adding statistics, quotations and cited sources, never the keyword stuffing it found made things worse. Where the deciding page is someone else's, Outreach drafts the pitch to it. Workspaces start in Draft, so every change waits in your inbox until you approve it, and anything that states a claim waits for your yes in every mode. After the change ships, the answers are re-read at 7, 14 and 30 days and the Analyst reports what moved, with the sample size.

What objections come up about GEO?

"GEO is just SEO rebranded." The tactics overlap, and the crawling and indexing basics are the same. What GEO adds is measured evidence about generated answers, including the finding that a classic tactic, keyword stuffing, made things worse.

"A single paper is not enough to act on." Agreed, which is why the useful response is to test it on your own questions rather than adopt it as a rule. The edits it recommends, sourced numbers, named sources and quotations, make a page better for people regardless.

"Our competitors already rank first, so GEO will not help us." The paper's result for fifth-ranked sites suggests rank in the retrieved list is not destiny inside an answer. Measure your own questions before deciding.

"We can add GEO markup and be done." There is no GEO markup. Google says no special structured data or AI file is needed for its AI features. GEO is an editorial practice, not a tag.