guide // how to // 2026
What is GEO? Generative engine optimization, the work of getting into AI answers
GEO, generative engine optimization, is the work of getting a brand's pages used, cited and recommended in the answers AI engines write: ChatGPT, Gemini, Perplexity, Claude, and Google's AI Overviews and AI Mode. The name comes from a paper written in 2023 by Pranjal Aggarwal and co-authors and presented at KDD 2024[1]. Most people use GEO for the same work as AEO, answer engine optimization, and Wikipedia lists LLMO, AIO and AI SEO as further names for it[2]. In practice it is two jobs: make your own pages easy for an engine to find, read and quote, and get your brand named on the other pages engines read. supereo is our pick, and we make it: an AI agent that does both jobs for you, then measures each change against control prompts it couldn't move. In the real run below, the client called one of supereo's tools: audit_page.
Updated . Prices checked 19 November 2026.
- The verdict
- The steps
- Real runs
- What the GEO paper actually measured
- What a later benchmark found in conversational search
- Why a tactic can work on one engine and fail on another
- What a GEO change looks like on a real page
- How to tell whether a GEO change worked
- Have it done: supereo
- Questions
- How we compiled this
- Sources
The verdict
GEO done for you, and proven
supereoour pick, and we make it
- best for teams that want AEO done, not a dashboard: supereo does everything it takes to get you recommended, from your site and pages to the roundups, placements and videos AI reads;
- it proves each move against control prompts, with sample sizes and ranges on every number;
- run it from Claude or ChatGPT: an MCP server with the same tools supereo's own agent uses.
From $149 a month
Run the free auditFree, no card, every answer shown.
The steps
Read what each engine says about your buyers' questions
Before changing anything, ask the engines the questions your buyers ask and read whole answers: which pages they cite, which brands they name, what they get wrong about you. Run the free audit to see 20 buyer prompts answered on five engines, with no card.
Let each engine's search crawler in
An engine can only quote a page its crawler may read. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers[29]. Anthropic says blocking Claude-SearchBot may reduce your site's visibility in Claude's search results[30]. Check robots.txt for both before you write a word.
Put the answer where a crawler meets it first
Open the page with a plain, one-sentence answer to the question it targets, in the HTML itself, ahead of menus and widgets. Don't park that answer behind a tab or an accordion: Microsoft warns that AI systems may not render hidden content, so key details can be skipped[32].
Back every claim with its source
In the GEO paper's own engine, citing sources, adding quotations and adding statistics did best of the methods it compared[9]. Yet in C-SEO Bench, the statistics method lowered rankings in most of the settings it measured[21]. Add a figure because it's true and sourced, never to look rigorous.
Get named on the pages engines read about your market
Answers aren't built from your site alone. Microsoft's guide to GEO says AI assistants reason over data on and off your site before deciding whether a brand is worth recommending[34]. Find the lists, reviews, threads and videos the engines cite for your buyers' questions, and earn a place on them with facts their authors can check.
Prove each change against prompts it couldn't move
Ask the target prompts and a set of control prompts again and again, before and after the change, and credit the change only with the gap between them. Change one thing on a page at a time, so the gap has one cause. The last section below explains why, and How supereo measures shows how supereo does it.
Real runs
a real run // 19 November 2026
Wikipedia's article on generative engine optimization, as a crawler reads it
Recorded from supereo's MCP server on its staging environment, https://staging.supereo.app/api/mcp. 1 call.
audit_page · 399 ms ·
Sent
{
"url": "https://en.wikipedia.org/wiki/Generative_engine_optimization",
"query": "what is geo"
}Received
{
"url": "https://en.wikipedia.org/wiki/Generative_engine_optimization",
"finalUrl": "https://en.wikipedia.org/wiki/Generative_engine_optimization",
"status": 200,
"ms": 18,
"signals": {
"title": "Generative engine optimization - Wikipedia",
"titleYear": null,
"h1": "Generative engine optimization",
"h1Count": 1,
"words": 2259,
"jsShell": false,
"h2": 4,
"h3": 0,
"questionHeadings": 0,
"listItems": 574,
"orderedListItems": 28,
"tables": 4,
"externalLinks": 35,
"dollarFigures": 3,
"percentFigures": 0,
"bestForLabels": 0,
"structuredData": [
"Article",
"Organization",
"ImageObject"
],
"author": "Contributors to Wikimedia projects",
"newestDate": {
"date": "2026-11-17T20:11:58.000Z",
"source": "last_modified",
"ageDays": 1
},
"noindex": false,
"canonical": "https://en.wikipedia.org/wiki/Generative_engine_optimization",
"queryInTitle": 0,
"queryInOpening": 1,
"openingText": "23 languages Български বাংলা Bosanski Català Čeština Deutsch Ελληνικά Español Eesti فارسی Français עברית Bahasa Indonesia Italiano 日本語 한국어 Latviešu Nederlands Português Română Русский Slovenščina 中文 E",
"aiCrawlersBlocked": [],
"llmsTxtServed": false
},
"issues": []
}audit_page read Wikipedia's article as a crawler without JavaScript gets it and checked it against the question what is GEO: its title and opening, its headings, lists and tables, its dates and its structured data.
What the GEO paper actually measured
The GEO paper didn't measure ChatGPT or Google's AI Overviews. Its authors built their own generative engine in two steps: fetch the top results for a query from Google's search engine, then have a language model write an answer from them[5]. They ran it over GEO-bench, a benchmark of 10,000 queries from diverse domains and sources[6].
Visibility needed a new yardstick, because an answer written from several sources has no list to rank in. The paper scored each source two ways: Position-Adjusted Word Count, which weighs how much of the answer comes from sentences citing the source and where that citation sits, and Subjective Impression, an overall score built from several judgments about the citation[7].
The authors tried nine ways of rewriting a source and compared each with the unchanged source[8]. Citing sources, adding quotations and adding statistics came out on top[9]. Keyword stuffing, the old SEO habit, gave little to no improvement in the engine's answers[10].
Two details limit how far that headline travels. The gains landed unevenly: with Cite Sources, the site ranked fifth in the search results gained visibility, while on average the top-ranked site lost some[11]. And the check on a live engine, Perplexity, didn't let Perplexity search the web: the authors uploaded the sources as files and used 200 samples from their test set[12].
What to take from it: in one controlled engine, how a source is written changed how much of the answer it shaped. That doesn't show the same edits move ChatGPT, Gemini or Google's AI Overviews on your buyers' questions, which is why the last section is about measuring on those engines.
What a later benchmark found in conversational search
C-SEO Bench, accepted at NeurIPS in 2025[14], set out to fix a gap in earlier work. Those evaluations looked at one web document adopting a method, while in reality many sites are likely to adopt the same tactics[15]. It covered two tasks, answering questions and recommending products, with three domains each[16].
Its finding was blunt: most published methods were largely ineffective and often pushed a document down the answer, while traditional SEO, getting the document ranked higher in what the model reads, worked far better[17]. Of 54 cases, only three showed a statistically significant gain[18].
It also measured what happens when everyone does it. Early adopters gained, then the gains fell steadily as more sites adopted the same method, which the authors call a congested, zero-sum game[19]. A GEO trick anyone can copy is worth less each time a competitor copies it. Facts, sources and a page that answers the question better than the rest are much harder to copy.
Read side by side, the two papers fit together. Wording helps a page that an engine already reads, and it rarely puts a page in front of the engine in the first place. So GEO work starts with being found and read, and only then turns to how each page is written.
Why a tactic can work on one engine and fail on another
Each engine finds its sources its own way. Google's AI features ground their answers in pages that its core ranking systems retrieve from the Search index[22]. ChatGPT search sometimes partners with other search providers and rewrites your question into queries it sends them[28]. Even within Google, AI Mode and AI Overviews may use different models and show different links[23].
Google's AI may also split one question into several searches. For someone asking how to fix a lawn full of weeds, its fan-out queries might cover herbicides, weeding without chemicals and keeping weeds away[24]. A page that answers one of those smaller questions well has its own way into the answer.
The research points the same way. In the GEO paper, how well each method worked varied across domains[13]. In C-SEO Bench, the rare methods that helped did so only for specific domains, tasks and models, and none helped on question answering or on Claude's Haiku model[20].
So treat a GEO tactic as a guess to check on each engine, not a rule. Ask the same prompts on every engine your buyers use and read the results engine by engine. supereo asks the same six engines (ChatGPT, Gemini, Perplexity, Claude, Google AI Mode and AI Overviews) for every brand it tracks, three times a week.
What a GEO change looks like on a real page
The real run on this page reads Wikipedia's article on generative engine optimization as a crawler without JavaScript gets it, checked against the question what is GEO. We chose it because it is a plain definition page for exactly this question. Its text is all in the HTML, so a crawler that runs no scripts still gets the definition. Its title spells out the practice but never says GEO, its headings ask no questions, and the opening text the run took is the article's language menu, not its first sentence.
A GEO change works on exactly those parts. Lead with a one-sentence answer that a crawler reaches before any menu: on a definition page, the definition itself, not a language list or a cookie notice. Put the question's own words in the title and the main heading, which AI systems read with the description to work out what a page is for[33]. Keep the answer in the HTML rather than behind a tab, since AI systems may not render hidden content[32].
The run also shows what the page gets by without. It serves no llms.txt, and Google says its search doesn't use such files[25]. What it does have is a recent edit, about a day old when the run read it, and freshness counts for more in an answer than in a ranking: Bing says that in grounding, a stale fact produces a misleading response[35].
How to tell whether a GEO change worked
A before and after on one answer proves little, because AI answers move on their own. When SE Ranking compared three sets of results for the same queries in Google's AI Mode, it often found completely different pages each time[37]. And a handful of answers can only separate big differences: Graphite showed it takes far more answers to tell two close rates apart[38].
The fix is a control. Pick the prompts a change targets and prompts it can't touch, ask both again and again, and credit the change only with how much more the targets moved. supereo grades each change that way: how much its target prompts moved, minus how much your other prompts moved, on each AI against its own before[39]. Those other prompts soak up whatever moved by itself, such as a model update or a competitor's launch[40].
Decide what counts before you start. An answer can mention your brand, recommend it, or cite one of your pages, and a GEO change can move one of those and leave the others where they were. Pick the one the change is meant to move, judge it on that, and keep the other two in view.
Give it time. A result on one of supereo's changes takes 14 to 60 days, and it reads worked, no change, made it worse or too close to call, with sample sizes and a range on every number. How supereo measures sets out each rule.
Have it done: supereo
01supereoour pick, and we make it
- best for teams that want AEO done, not a dashboard: supereo does everything it takes to get you recommended, from your site and pages to the roundups, placements and videos AI reads;
- it proves each move against control prompts, with sample sizes and ranges on every number;
- run it from Claude or ChatGPT: an MCP server with the same tools supereo's own agent uses.
From $149 a month
- Does the work
supereo writes those pages for you, aimed at the questions where AI skips you.
supereo.app - Proves it against controls
Each change has a few target questions, and we compare them with your other questions, the ones no change targets.
supereo.app - Runs from Claude or ChatGPT
Run the same loop from Claude, Cursor or Codex. Your agent gets every tool supereo's own agent uses: your buyers' questions, every answer, the pages AI cites, your brand facts.
supereo.app - Free audit
free audit // about 2 minutes // no card
supereo.app
| Plan | A month | Prompts | AI engines |
|---|---|---|---|
| Autopilot | $149 | 40 | ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Claude |
Prices from supereo.app, checked 19 November 2026. What it does, from its own pages, checked 16 November 2026.
How the proof works, with every number it runs on: How supereo measures.
Questions
What is GEO?
Generative engine optimization: getting your brand's pages used, cited and recommended when AI engines write answers. The paper that named it described GEO as a way to help content creators improve their content's visibility in generative engine responses[4].
What name is GEO short for?
In marketing, generative engine optimization. The same letters mean other things too: in spaceflight, GEO is a geostationary orbit, where satellites fly above the equator matching Earth's rotation[36]. If you met GEO in a marketing brief, it means the AI search work this page describes.
Is GEO different from SEO?
In its target more than its methods: SEO earns a place in a list of links, GEO a place inside the answer. Google says that from its side, optimizing for generative AI search is still SEO[26]. C-SEO Bench went further and found traditional SEO beat most GEO-specific content methods[17]. GEO vs SEO vs AEO compares them pair by pair.
How does AEO work?
Like GEO, under another name. An answer engine searches the web, picks trusted sources and writes one answer from them[31], so AEO is the work of getting your pages among those sources and your brand into the answer. What is AEO walks through it.
Is GEO the same as AEO?
For most people, yes. Google describes both as terms for work aimed at visibility in AI search[26], and Wikipedia notes the terms overlap with no consensus definition[3]. Where some draw a line between them is in GEO vs SEO vs AEO.
Can I do Google SEO myself?
Often, yes: Google says that if you run a small local business, you can probably do much of the work yourself[27]. GEO's basics are the same kind of work: open your site to search crawlers, answer questions plainly, keep facts current. The slow parts are earning mentions on other sites and proving what each change did.
Does GEO work?
Some of it, under some conditions. The GEO paper's best methods raised visibility in its own engine[9], while C-SEO Bench found most methods largely ineffective in conversational search[17]. Every published result on whether it works is collected with its sample size.
What is a generative engine?
A search engine that writes the answer itself. It retrieves pages, then a language model writes one response from them, which is how the GEO paper built its own engine[5]. Google calls the same idea grounding when its AI features answer from pages in its Search index[22].
What are the best GEO tools?
The best GEO tools compares them on four jobs, from doing the work to proving it against controls, each read from the tool's own pages, with supereo as our pick.
How we compiled this
Where this page's words say how an AI engine finds, reads or cites pages, the sentence links the engine's own documentation, or another source, and the day it was read.
The runs are real calls to supereo's MCP server on its staging environment, made as a house account, recorded as they happened: the arguments as sent, each answer as the client received it, its JSON indented and trimmed only where it says so, the time each call took and when it ran.
We make supereo, so it is the one tool this page names as a pick.
How supereo measures its own results, with every number it runs on: How supereo measures.
supereo (supereo.app) is an AI agent that does everything it takes to get your brand recommended by ChatGPT, Gemini, Perplexity, Claude and Google, on your site and off it: site fixes, new pages, roundups, placements and video. It proves what worked against control prompts. Made by superagnt.
Sources
The term "generative engine optimization" was introduced in a 2023 paper by Pranjal Aggarwal and co-authors, later presented at the KDD 2024 conference.
en.wikipedia.org, read 19 November 2026.Other terms for the same concept include answer engine optimization (AEO), large language model optimization (LLMO), artificial intelligence optimization (AIO), and AI SEO.
en.wikipedia.org, read 19 November 2026.Several overlapping terms describe related practices, and usage varies across practitioners, vendors, and publications, with no consensus definition.
en.wikipedia.org, read 19 November 2026.To address this, we introduce Generative Engine Optimization (GEO), the first novel paradigm to aid content creators in improving their content visibility in generative engine responses through a flexible black-box optimization framework for optimizing and defining visibility metrics.
arxiv.org, read 19 November 2026.The first step involves fetching relevant sources for input query, followed by a second step where an LLM generates a response based on the fetched sources. Similar to prevous works, we do not use summarization and provide the whole response for each source. Due to context length limitations and quadratic scaling cost based on the context size of transformer models, only the top 5 sources are fetched from the Google search engine for every query.
arxiv.org, read 19 November 2026.To facilitate faithful and extensive evaluation of GEO methods, we propose GEO-bench , a benchmark consisting of 10000 queries from diverse domains and sources, adapted for generative engines.
arxiv.org, read 19 November 2026.Performance was measured using two metrics: Position-Adjusted Word Count and Subjective Impression . The former considers word count and citation position in the GE’s response, while the latter computes multiple subjective factors, giving an overall impression score.
arxiv.org, read 19 November 2026.We evaluate 9 different proposed GEO methods as described in Section 2.2.2 . We compare them with a baseline, which measures the impression metric of unmodified website sources.
arxiv.org, read 19 November 2026.Specifically, our top-performing methods, Cite Sources, Quotation Addition, and Statistics Addition, achieved a relative improvement of 30-40% on the Position-Adjusted Word Count metric and 15-30% on the Subjective Impression metric.
arxiv.org, read 19 November 2026.Finally, we evaluate keyword stuffing, i.e., adding more relevant keywords to website content. While widely used for Search Engine Optimization, we find such methods offer little to no improvement on generative engine’s responses.
arxiv.org, read 19 November 2026.For example, the Cite Sources method led to a substantial 115.1% increase in visibility for websites ranked fifth in SERP, while on average, the visibility of the top-ranked website decreased by 30.3%.
arxiv.org, read 19 November 2026.Since perplexity.ai does not allow the user to specify source URLs, we instead provide source text as file uploads to perplexity.ai while ensuring all answers are generated only using the file sources provided. We evaluate all our methods on a subset of 200 samples of our test set.
arxiv.org, read 19 November 2026.Moreover, we show the efficacy of these strategies varies across domains, underscoring the need for domain-specific optimization methods.
arxiv.org, read 19 November 2026.Accepted at NeurIPS Datasets & Benchmarks 2025
arxiv.org, read 19 November 2026.Moreover, existing evaluations consider only a single-actor scenario where only one web document adopts a C-SEO method; in reality, multiple players are likely to competitively adopt the cutting-edge C-SEO techniques, drawing an analogy from the dynamics we have seen in SEO.
arxiv.org, read 19 November 2026.We consider two search tasks, question answering and product recommendation, with three domains each.
arxiv.org, read 19 November 2026.Our experiments reveal that most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking, which is opposite to what is expected. Instead, traditional SEO strategies, those aiming to improve the ranking of the source in the LLM context, are significantly more effective.
arxiv.org, read 19 November 2026.Out of 54 cases, we uncover only three where the ranking improvements are statistically significant.
arxiv.org, read 19 November 2026.We observe that while early adopters experience significant gains, these gains decrease steadily as more players adopt the same method. This pattern reveals two key characteristics of C-SEO: it is both a congested and a zero-sum game.
arxiv.org, read 19 November 2026.We found no method effective for the question answering task, and no methods effective for the Haiku 3.5 model (results in appendix). These findings highlight the very limited effectiveness of most stylistic or content-editing strategies: when a C-SEO method significantly boosts the ranking, it does so only for specific domains, tasks, and models.
arxiv.org, read 19 November 2026.This is not a marginal phenomenon: for example, the Statistics method decreases rankings in 19 out of 24 evaluated settings.
arxiv.org, read 19 November 2026.Retrieval-augmented generation (RAG) : A technique (also known as grounding) used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index.
developers.google.com, read 19 November 2026.AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.
developers.google.com, read 19 November 2026.Query fan-out : A set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query. For example, if the original user's query is "how to fix a lawn that's full of weeds", fanout queries might include "best herbicides for lawns", "remove weeds without chemicals", and "how to prevent weeds in lawn".
developers.google.com, read 19 November 2026.LLMS.txt files and other "special" markup : You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them.
developers.google.com, read 19 November 2026."AEO" stands for "answer engine optimization" and "GEO" for "generative engine optimization". These are both terms you may see used to describe work specifically focused on improving visibility in AI search experiences. From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.
developers.google.com, read 19 November 2026.If you run a small local business, you can probably do much of the work yourself.
developers.google.com, read 19 November 2026.To provide relevant responses to your questions, ChatGPT search sometimes partners with other search providers. When it does, ChatGPT search typically rewrites your query into one or more targeted queries that it sends those providers.
help.openai.com, read 19 November 2026.OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features. Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.
developers.openai.com, read 19 November 2026.Disabling Claude-SearchBot on your site prevents our system from indexing your content for search optimization, which may reduce your site's visibility and accuracy in user search results.
support.claude.com, read 19 November 2026.An answer engine is a tool designed to give you direct, detailed answers to your questions. Perplexity serves as an answer engine by searching the web, identifying trusted sources, and synthesizing information into clear, up-to-date responses.
perplexity.ai, read 19 November 2026.Your page title, description, and H1 tag (the top-level HTML heading) are important signals AI systems use to interpret purpose and scope.
about.ads.microsoft.com, read 19 November 2026.AI assistants, browsers, and agents don't just index your site. They reason over data on and off your site and decide whether your brand is worth recommending.
about.ads.microsoft.com, read 19 November 2026.In search, stale content degrades ranking. In grounding, a stale fact produces a misleading response.
blogs.bing.com, read 19 November 2026.Satellites in geostationary orbit (GEO) fly above Earth’s equator, moving from west to east, exactly matching Earth’s rotation
esa.int, read 19 November 2026.The average overlap of exact URLs between the three datasets was just 9.2%. This means that AI Mode often chooses completely different pages each time, even when the query doesn’t change.
seranking.com, read 19 November 2026.For example, with 10 responses, we can statistically distinguish between 75% and 25%, but not 75% and 50%. With 50 responses, we can statistically distinguish between 50% and 30%, but not 50% and 40%.
graphite.io, read 19 November 2026.Added by the change : how much the target prompts moved, minus how much your other prompts moved, on each AI against its own before, then averaged over the AIs, each counting the same.
supereo.app, read 19 November 2026.Your other prompts : every prompt you track that doesn't name you and that no change targets, pooled each time the test is graded, none whose answers cite the page. They show how AI answers moved on their own: a model update, a competitor's launch, a slow week.
supereo.app, read 19 November 2026.