Generative Engine Optimization (GEO): what it is and how it works

Generative engine optimization (GEO) is the work of getting a brand mentioned, described correctly and cited as a source in answers written by AI systems such as ChatGPT, Gemini, Perplexity and Google's AI Overviews. It builds on SEO, because most of these systems look up web pages before they answer, and adds prompt-based measurement and content written to be quoted.

Key takeaways

  • The term GEO comes from a paper by researchers at Princeton University and IIT Delhi with two independent co-authors, posted in November 2023 and presented at KDD 2024.
  • The often quoted 'up to 40%' is the 41% gain of a single method, Quotation Addition, on one measure, Position-Adjusted Word Count, in the paper's own benchmark.
  • Google says its AI features rest on its core Search ranking and quality systems and need no llms.txt file, no chunking and no special schema markup.
  • In SparkToro's 2026 study, there was less than a 1 in 100 chance that ChatGPT or Google's AI gave the same list of brands in two responses to the same prompt.
  • A 2026 research survey found that no reviewed GEO technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability.

What is generative engine optimization?

Generative engine optimization (GEO) is the work of getting a brand mentioned, described correctly and cited as a source in answers that AI systems write, such as ChatGPT, Gemini, Perplexity and Google's AI Overviews and AI Mode. It builds on search engine optimization, because most of these systems look up web pages before they answer, and adds measurement and content aimed at the answer itself.

A classic search engine returns a ranked list of links. A generative engine writes a response in prose and links some of the pages it drew on. In a list you want a high position and a click. In a written answer you want your brand to appear, the description of it to be correct, and one of your pages among the linked sources.

The audience asking these systems is large. In a Pew Research Center survey from February 2026, 49% of US adults said they use AI chatbots, compared with 33% in 2024 on a slightly different question [13].

Where does the term GEO come from?

The term comes from a research paper titled "GEO: Generative Engine Optimization", first posted on arXiv on 16 November 2023 and accepted at the KDD 2024 conference in Barcelona [1]. Three of its six authors are at Princeton University, one is at IIT Delhi, and two list independent affiliations [1].

The authors built GEO-bench, a benchmark of 10,000 queries collected from nine sources [1]. They then tested nine ways of rewriting a web page: Authoritative, Statistics Addition, Keyword Stuffing, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words and Technical Terms [1]. Their main test engine was a setup of their own, which fetched the top five Google results for each query and had gpt-3.5-turbo write the answer [1].

The result most people repeat is "up to 40%" [1]. The single best method, Quotation Addition, improved a page's visibility by 41% over the unmodified page on one measure, Position-Adjusted Word Count, which scores how prominently a source features in the generated answer [1]. That figure is a benchmark result for one method on the paper's own test set. It is no combined effect of several methods, and it says nothing about clicks or traffic.

The paper names Cite Sources, Quotation Addition and Statistics Addition as its top-performing methods, with relative improvements of 30 to 40% on Position-Adjusted Word Count and 15 to 30% on a second measure, Subjective Impression [1]. Keyword stuffing, a familiar SEO tactic, gave "little to no improvement" and scored slightly below the baseline [1]. Lower-ranked pages gained most: in one experiment, Cite Sources raised the visibility of the site ranked fifth in Google by 115.1%, while the top-ranked site lost 30.3% on average [1].

The authors also ran a smaller test on Perplexity.ai, a deployed engine, with 200 samples. Quotation Addition came first again, with a 22% improvement on Position-Adjusted Word Count, and keyword stuffing did 10% worse than the baseline [1]. Neither engine in the study was ChatGPT or one of Google's AI features. The paper is still the reference point for the field: in our prompt tracking from November 2026, its arXiv page is one of the five URLs with the largest Google traffic share for this topic.

How generative engines find and choose sources

Most engines that answer questions about products and brands search before they write. They split the question into searches, retrieve pages from an index and write the answer from what they retrieved, linking some of those pages.

Google documents this for its own features. AI Overviews and AI Mode use retrieval-augmented generation, also called grounding, which relies on Google's core Search ranking systems to retrieve relevant pages from its Search index [2]. Both features "may use a 'query fan-out' technique" [3], which Google's guide describes as a set of concurrent, related queries that the model generates to fetch additional results [2]. Google does not publish how the model then picks which of the retrieved pages it quotes.

Evergreen Media, an SEO agency, sorts generative engines into three types [9]. It is one publisher's classification, and a useful one:

TypeExamples given by Evergreen MediaWhere the answer comes from
Search-basedGoogle AI Overviews, AI Mode, PerplexityReal-time web indexing
HybridGoogle Gemini, ChatGPT SearchTraining data combined with live web content
Training-basedModels answering without web accessWhat the model learned during training

For the training-based case, Evergreen's advice is that influence comes "only through long-term efforts like digital PR and building a broader digital footprint" [9].

Each provider runs its own crawlers, and most separate the crawler for search answers from the one that collects training data:

  • OpenAI: OAI-SearchBot serves ChatGPT search. Sites that block it "will not be shown in ChatGPT search answers, though can still appear as navigational links" [6]. GPTBot collects training data, and the two settings are independent [6].
  • Perplexity: PerplexityBot exists to surface and link websites in Perplexity's results and is not used to crawl content for foundation models. Perplexity recommends allowing it in robots.txt [7].
  • Anthropic: Claude-SearchBot serves search answers, ClaudeBot collects training data and Claude-User fetches pages when a user asks. Disabling Claude-SearchBot "may reduce your site's visibility and accuracy in user search results" [8].

So a site can opt out of training and still be eligible for search answers. A robots.txt file that blocks every AI crawler by name blocks both.

GEO, SEO, AEO and AI SEO: what is the difference?

GEO, AI SEO, answer engine optimization (AEO), LLMO and AIO are used today as largely interchangeable names for optimizing for AI-written answers, and Wikipedia's entry notes that usage varies "with no consensus definition" [10]. SEO is the older discipline, aimed at the ranked list of results, and GEO depends on it.

SEOGEO (also called AEO, AI SEO)
Where you appearA ranked list of linksAn answer written by an AI system
What counts as successPosition and clicksBeing named, described correctly and cited
Unit of researchKeywordPrompt, usually a full question with context
How you measureSearch Console, rank trackingA fixed prompt set, run repeatedly in several AI systems
FoundationIndexing, helpful content, linksThe same foundation, plus the sources the AI systems cite

The unit of research matters most in practice. People type longer questions when they expect an answer, and longer queries trigger AI answers more often. In Pew's March 2025 data, 8% of one- or two-word Google searches produced an AI summary, against 53% of searches with ten words or more [12]. The full comparison, including what changes in budgets and reporting, is in GEO vs SEO.

What Google says about optimizing for its AI features

Google's position, set out in a guide last updated in July 2026, is that ordinary SEO remains the basis: "The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems" [2]. A second Google page states: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" [3]. A page has to be indexed and eligible to be shown in Search with a snippet [3].

On what helps most, Google says unique, valuable content "will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide" [2]. It also advises: "Don't just recycle what others on the internet have already said, or could easily be produced by a generative AI model" [2].

The same guide lists what you can skip, in Google's words [2]:

  • "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." This covers llms.txt.
  • "There's no requirement to break your content into tiny pieces for AI to better understand it."
  • "You don't need to write in a specific way just for generative AI search."
  • "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add." Google still recommends structured data as part of overall SEO.

Google addresses the GEO label directly: "While terms like Answer Engine Optimization (AEO) or Generative Engine Optimization (GEO) are common online, many suggested "hacks" aren't effective or supported by how Google Search actually works" [2]. On chasing mentions on other sites, it says that "seeking inauthentic "mentions" across the web isn't as helpful as it might seem" [2].

Two warnings concern volume. Creating separate content for every variation of a query or every fan-out query is covered by this sentence: "doing so primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy" [2]. That policy defines scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users" [5]. Google's guidance on AI-written content adds: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing" [4].

All of this applies to Google Search; for OpenAI, Perplexity and Anthropic we found crawler documentation and no comparable guide. In our prompt tracking from November 2026, Google's guide was among the pages cited by Gemini, Google AI Mode, Google AI Overviews and Perplexity when they answered questions about GEO.

Which GEO measures have evidence?

Few GEO measures rest on more than one study or one provider's statement. Crawler access and indexing are documented preconditions; the content techniques have benchmark evidence from a single 2023 paper; and several popular tactics are contradicted by Google's own statements.

MeasureEvidenceStrength
Be indexed and eligible for a snippet in GoogleGoogle's documentation [3]Documented precondition for Google's AI features
Allow the search crawlers of OpenAI, Perplexity and AnthropicProvider documentation [6][7][8]Documented precondition for their search answers
Publish unique, valuable contentGoogle's guide [2]Google's own statement, with no published effect size
Add quotations, statistics and cited sourcesThe GEO paper [1]One benchmark study, a 2023 test engine, plus a 200-sample Perplexity check
Keyword stuffingThe GEO paper [1]Benchmark result: no gain, slightly worse than the original page
llms.txt, chunking, special schema for AIGoogle's guide [2]Google says its Search does not use them; we found no statement from other providers
Inauthentic mentions on other sitesGoogle's guide [2]"Not as helpful as it might seem", per Google
A separate page for every query variationGoogle's guide [2]A violation of the scaled content abuse policy if done primarily to manipulate

The paper also found that the best method differs by subject area, and its authors advise that "Website-owners can choose relevant GEO strategy based on their target domain" [1]. Treat its techniques as hypotheses to test on your own prompts.

Example: one paragraph made citable

What follows is an example written for this article. The company is fictional and every figure in it is invented for illustration.

Before:

We are a leading provider of innovative door solutions for the logistics industry. With decades of experience and a passion for quality, we deliver products that meet the highest standards and help our customers succeed.

After:

High-speed roll doors for freezer rooms keep cold air in by closing within seconds after each forklift passes. Our freezer model is rated for minus 30 °C and closes in 1.5 seconds, according to our March 2026 test report (linked). It pays off at around 40 or more passes per hour; below that, an insulated sliding door is cheaper to run.

The first version could describe any company and gives an AI system nothing to quote. The second answers a buyer's question in its first sentence, gives figures with a dated source, and says when the product is the wrong choice. Statistics and cited sources are two of the techniques the GEO paper tested [1]; the statement of fit tells an AI system whom the product suits, which it needs before it can recommend anything.

Finding the questions buyers ask an AI

A GEO program needs a prompt set built from the questions buyers actually put to an AI system. Search data shows what people type into Google, in volume, and misses most of the longer, situational questions people ask a chatbot.

Digital Twins of the audience, powered by neurotwins by neuroflash, phrased questions like these about GEO itself:

  • "how does generative engine optimization work with existing seo practices"
  • "what are generative engine optimization strategies for b2b clients"
  • "what are the key challenges in tracking generative engine optimization visibility"

Each asks for something a keyword list would not reveal: how GEO fits work already running, what applies in B2B, and what makes tracking hard.

At geoagents.io, agents build the prompt set from both sources: clustered search and People-Also-Ask data from KeywordCluster for what people search, and Digital Twins of the audience for what buyers would ask an AI. The agents then track the prompts, find the gaps and draft the content, and a person on the brand's side releases every piece. How GEO agents work describes each step, and GEO with agents describes the service. Paid placements inside AI answers are a separate discipline, covered in Generative Engine Advertising.

A 90-day GEO plan

A first GEO quarter has four phases. Measure where you stand, fix access and indexing, rework the pages behind your most important prompts, and measure again. The plan below is our recommendation. The phases follow from the evidence above, and the week ranges are a practical rhythm that no study prescribes.

WeeksWhat to doWhat you have at the end
1 to 2Build a prompt set from search data and buyer questions, grouped by topic and intent. Run every prompt several times in each AI system you care about.A baseline: mention rate, citation rate and the errors AI systems make about you
3 to 4List the URLs the AI systems cite for your prompts. Check indexing, snippet eligibility and robots.txt rules for each provider's search crawler.A source map and a list of technical fixes
5 to 8Rework the pages for your most important topics: direct answer first, sourced statistics and quotations, your own data, clear statements of fit. A person reviews every page before it goes live.Pages built to be quoted
9 to 12Correct wrong or outdated facts about you on the third-party pages the AI systems cite, where you can. Re-run the full prompt set.A second measurement to compare with the baseline

Keep the prompt set fixed across the quarter. If you change the prompts, you can no longer tell a real change from a change in what you asked.

How do you measure GEO?

You measure GEO by running a fixed set of prompts repeatedly in several AI systems and counting, per answer, whether your brand is mentioned, whether one of your pages is cited, whether the facts are correct and how the brand is described. Compare each of these with your main competitors on the same prompts.

Repetition is not optional. In a 2026 study by SparkToro with the tracking company Gumshoe.ai, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI Overviews and AI Mode, 2,961 runs in total [11]. The finding: "there's a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses" [11]. The author calls the study small and discloses a conflict of interest with Gumshoe. His recommendation: "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric" [11].

Measure each engine on its own. BrightEdge reports that the top 100 cited sources of any two engines overlap by only 16% to 59%, across ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews [15]. A page that one engine cites often can be absent from another.

For Google, there is now first-party data. Google launched Search Generative AI performance reports in Search Console on 3 June 2026 and rolled them out to all websites worldwide on 31 August 2026 [14]. They show how often your URLs appeared in AI Overviews, AI Mode and generative features in Discover [14]. The full method, including sample sizes and margins of error, is in How to measure AI visibility.

Where GEO falls short

GEO is a young practice with thin evidence. A critical survey of GEO research from 2023 to 2026, published as a preprint in July 2026, concludes that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability" [16]. The author describes GEO as "a stochastic, partially observable pipeline" [16]. The survey has not been peer reviewed, and it matches what the primary sources show.

The limits worth planning around:

  • The founding numbers come from a lab. The GEO paper tested a 2023 setup built on gpt-3.5-turbo, plus a 200-sample check on Perplexity.ai [1]. Today's ChatGPT, Gemini and Google features work differently, and nobody has published a replication on them in the sources we checked.
  • Nobody controls the answer. The model decides what it writes. Any provider who guarantees placement in an AI answer is promising something no one can deliver.
  • Answers vary. The same prompt returns different brands from run to run [11], so small changes between two measurements are often noise.
  • AI answers send fewer clicks. In Pew's US data from March 2025, users clicked a traditional search result on 8% of visits when an AI summary appeared and on 15% when none appeared, and they clicked a link inside the summary on 1% of visits [12]. Around one in five Google searches in that month produced an AI summary [12]. Being cited matters for what buyers read, and less for traffic. Pages that are indexed, open to search crawlers and answer the question with sourced facts are the pages AI systems can use. Measuring repeatedly is the only way to see whether they do.

Questions and answers

What is GEO and how does it work?

GEO is the work of getting a brand named, described correctly and cited in answers written by AI systems. Most of these systems search the web, retrieve pages and write an answer from them. GEO makes sure your pages can be retrieved, answer the question directly, carry sourced evidence, and that you measure the result with a fixed set of prompts run repeatedly.

How do I optimize for generative AI answers?

Keep pages indexable and let the search crawlers of each AI provider in. Open every section with a direct answer, add statistics and quotations with their sources, and publish information nobody else has. Then track a fixed prompt set in several AI systems over time and fix the pages where competitors appear and you do not.

Is AI SEO the same as generative engine optimization?

In current usage, mostly yes. GEO, AI SEO, answer engine optimization (AEO), LLMO and AIO are all used for optimizing for AI-written answers, and Wikipedia notes there is no consensus definition. Google uses the terms AEO and GEO too, and warns that many tactics sold under them do not work for Google Search.

Do I need an llms.txt file for AI search?

Not for Google. Google states that its Search does not use machine-readable AI text files such as llms.txt, special markup or Markdown copies of pages. We found no statement from other AI providers on llms.txt, so treat the file as a low priority next to indexable, well-sourced pages.

How does generative engine optimization work with existing SEO practices?

It sits on top of them. Google says its AI features rely on its core Search ranking and quality systems, so indexing, helpful content and a clean technical setup remain the base. GEO adds a prompt set as the unit of research, measurement of mentions and citations in AI answers, and content written to be quoted.

Sources

  1. 1
    GEO: Generative Engine OptimizationAggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande (arXiv, KDD 2024), accessed arxiv.org
  2. 2
    Optimizing your website for generative AI features on Google SearchGoogle Search Central, accessed developers.google.com
  3. 3
    AI features and your websiteGoogle Search Central, accessed developers.google.com
  4. 4
    Google Search's guidance on using generative AI content on your websiteGoogle Search Central, accessed developers.google.com
  5. 5
    Spam policies for Google web searchGoogle Search Central, accessed developers.google.com
  6. 6
    Overview of OpenAI CrawlersOpenAI, accessed platform.openai.com
  7. 7
    Perplexity CrawlersPerplexity, accessed docs.perplexity.ai
  8. 8
    Does Anthropic crawl data from the web, and how can site owners block the crawler?Anthropic, accessed support.claude.com
  9. 9
    Generative Engine Optimization (GEO) explainedEvergreen Media, accessed evergreen.media
  10. 10
    Generative engine optimizationWikipedia, accessed en.wikipedia.org
  11. 11
    NEW Research: AIs are highly inconsistent when recommending brands or productsRand Fishkin, SparkToro, accessed sparktoro.com
  12. 12
    Google users are less likely to click on links when an AI summary appears in the resultsPew Research Center, accessed pewresearch.org
  13. 13
    Americans and AI 2026: Chatbots, Smart Devices and Views on ImpactPew Research Center, accessed pewresearch.org
  14. 14
    Search Generative AI performance reportsGoogle Search Central Blog, accessed developers.google.com
  15. 15
    Why AI Engines Cite Different Sources but Recommend the Same BrandsBrightEdge, accessed brightedge.com
  16. 16

Powered by neurotwins, the Digital Twins by neuroflash

Search data by KeywordCluster

Buyer questions in this article come from Digital Twins of the audience and from search data.

Want your brand in these answers?

Book an intro call. We look at the prompts that matter in your market.

Book an intro call (opens in a new tab)