Updated on 7 November 2026, by João Leite, Founder of murmur.
Yes, there are scientific studies proving that GEO works: since 2023, they show that changing the shape of a page changes its space in the AI answer. GEO, an acronym created in one of those studies, is the work of making a company appear in the answers of ChatGPT and the other AI models. Olivier Martinez reviewed 45 studies published up to July 2026 and gave high confidence to this effect: changing a page that the AI has already found changes the position, the citation or the use it makes of that page.
- The study that created the term GEO measured the gain of each technique. With quoted sentences, the page’s share of the AI answer went from 19.3% to 27.2%, and with statistics, to 25.2%.
- Citing sources helped the trailing page more. On the page that ranked fifth among the five in the search, the space in the answer grew 115.1%.
- With the entire search in the test, rewriting title, description and headings moved the page up. In SAGEO Arena, with the structured data also rewritten, it ranked among the top 20 of the selection in 71% of the questions, against 58%.
- On a site with a comparison group, the changes brought more visits from ChatGPT. The authors estimate 1.82 times more traffic coming from ChatGPT in the part of the site that received the package of changes.
- murmur’s case studies show Brazilian brands gaining space. Visio, retail software, appeared in 2% of ChatGPT answers in June 2026 and reached 22% in July, with the same 135 questions in both months.
Since 2023, dozens of scientific studies have tested what changes a page’s space in the AI answer
There are already dozens of scientific studies on GEO, and each one tested a part of the path between the question and the AI answer. murmur, a GEO agency in Brazil that places brands in the answers of ChatGPT, Claude, Google AIO and other AI agents, counted 50 papers on the topic on arXiv on 3 October 2026. Judging by the phrase “generative engine optimization” in the abstract, the first is from November 2023, and 44 came out in 2026. arXiv is the open repository where researchers publish papers, many before any review. The studies read alongside client case studies are in the guide GEO works: what the studies and the case studies show.
On the path to the answer, the AI decides whether it will search, finds pages, chooses some, puts them in order, writes and cites. Laboratory tests hand the pages to the AI, or build the entire search, and measure what happens to each page up to the written answer. Studies with the traffic of a real site count the visits that arrive afterward.
Not every paper carries the same weight. A peer-reviewed one went through review by specialists before being published at a conference or in a journal, such as the paper that created the term GEO, at KDD 2024, and SAGEO Arena, at KDD 2026. A preprint is published by the authors themselves, still without review. The table says what each study in this guide tested, what type it is and what it showed.
| study | what it tested | under what condition | what it showed | peer-reviewed |
|---|---|---|---|---|
| Aggarwal and others, 2023 | rewriting one of five pages in nine ways | the five pages handed to GPT-3.5, across a thousand questions | citing sources, bringing in quoted sentences and adding statistics increased the page’s share of the answer | yes, KDD 2024 |
| SAGEO Arena, 2026 | rewriting the text or the title, description, headings and structured data fields | a search built by the authors, with real pages taken from Google | the four rewritten fields moved the page up in the search’s first selection | yes, KDD 2026 |
| Watanabe and Nakayashiki, 2026 | a package of changes on part of a site’s pages | traffic coming from ChatGPT, with the rest of the site as a comparison | 1.82 times more ChatGPT traffic in the part that changed, by the authors’ estimate | no, it is a preprint |
| Martinez, 2026 | the evidence of 45 studies published from November 2023 to July 2026 | a review that gives each claim in the field a degree of confidence | high confidence that changing a page the AI already has changes the use it makes of that page | no, it is a preprint |
The study that created the term GEO measured how much of the AI answer comes from each page
The paper that created the term GEO measured how much of the AI answer came from each page, before and after rewriting one of them. Pranjal Aggarwal and colleagues gave GPT-3.5, an OpenAI model, the first five Google pages for each question, and it wrote the answer citing them. One of the five was rewritten in nine ways, across the thousand questions in the test part of a bank of 10 thousand questions. The origin of the name is in the guide what is GEO.
Aggarwal and colleagues’ measure splits the answer among the five pages, by each one’s words and by the place where they appear. The parts add up to 100%, and without any change the chosen page held, on average, 19.3%. Since the sum is always 100%, the space the rewritten page gains comes from the other four, which answer the same question.
The authors point to three techniques as the ones with the best results. With sentences from reliable sources in quotes, the page’s share went from 19.3% to 27.2%. With statistics in place of a description without numbers, it went to 25.2%, and with the citation of reliable sources, to 24.6%.
The authors also tested Perplexity, a commercial AI search engine, on 200 questions, with the texts sent as a file. There, quoted sentences increased the page’s share by 22%.
The “up to 40%” of the 2023 paper comes from cited sources, quoted sentences and statistics
The “up to 40%” in the abstract of the 2023 paper is the relative gain of the three best-performing techniques. In the authors’ words:
“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.”
The first measure, the Position-Adjusted Word Count, is the page’s share of the answer, described above. The second, the Subjective Impression, is a score that an AI gives to the cited page, by criteria such as its relevance and its influence. The biggest gain was from quoted sentences, from 19.3% to 27.2%, about 41% more.
The authors also measured the gain by each page’s position in the search. With the three techniques, the page that ranked fifth among the five was the one that gained the most. On it, citing sources increased the space in the answer by 115.1%, quoted sentences by 99.7%, and statistics by 97.9%.
With the entire search in the test, rewriting title, description, headings and structured data moved the page up
In SAGEO Arena, which put the entire search in the test, rewriting the title, the description, the headings and the structured data moved the page up in the first selection. Sunghwan Kim and colleagues, from two universities in Seoul, built a search in three parts, in a study accepted at KDD 2026. A selection matches the words of the question and of the page, a filter reorders the candidates, and the AI writes citing the sources. The 2,700 questions, from nine areas, had on average 63 real candidate pages each, taken from Google.
With those four fields rewritten, the page ranked among the top 20 of the selection in 71% of the questions, on average, against 58% without the change, a rise of 22%. Structured data is the code that describes each piece of information on the page to machines, such as name, price and date.
The authors also record that putting the answer at the beginning of the document earned a higher score in the reordering, the step that chooses the pages the AI will read. In the test search, the first selection matches the words of the question with those of the page, and the person asking generally uses the common word. What this asks of the writer is in the guide how to write a text to appear in ChatGPT.
Studies with the traffic of real sites show more visits after the changes
The studies with the traffic of a real site, discussed in Martinez’s review, show more visits after the changes. Keisuke Watanabe and Kazuki Nakayashiki studied glasp.co, a site with hundreds of thousands of question-and-answer pages about YouTube videos. In January 2026, part of the site received a package of AEO changes, another name for the work of appearing in AI answers: duplicate URLs merged, titles in question form, rewritten summaries and new pages. The rest of the site served as a comparison.
Traffic coming from ChatGPT grew across glasp.co as a whole, so the authors compared the part that changed with the part that stayed the same. After discounting ChatGPT’s own growth, their model estimates 1.82 times more traffic in the part that received the package, with 95% confidence that the real value falls between 1.31 and 2.54 (the confidence interval). The number applies to the whole package, with the new pages included.
Pinterest, in another study with real traffic, reports 20% more organic traffic with the company’s own GEO system, compared with a control group. How to measure the brand in the answers, beyond the click, is in the guide about the reliable presence metric.
The review of 45 studies confirms that changing the page changes the use the AI makes of it
Olivier Martinez’s review, with 45 studies published from November 2023 to July 2026, confirms that changing a page the AI already has changes the use it makes of that page. Martinez gives each claim in the field a confidence level, and this one came out with high confidence:
“A document already placed in the context can causally alter its rank, citation, or use.”
In the same review, Martinez rates the support the studies give to each factor of the page. The page’s relevance to the question has strong support in a controlled setting, and the evidence that can be taken from the text, such as the number with its source, has moderate to strong support. Recency, prices and dates have moderate support. The effect of the same change varies across AI models, which is why murmur measures each brand in each of them.
How to read a scientific study’s percentage before using it
Before using a study’s percentage, check five things: what was measured, the AI, the repetitions, the comparison and who did it.
- What was measured. With pages handed to the AI, the study measures the space in the answer, and with a site’s traffic, it measures the visits.
- The AI and how it is accessed. An API that receives the URLs, a model with a web search tool and the interface the public uses are different systems, Martinez warns.
- The repetitions. The AI answer varies, and Martinez asks for several repetitions, on different dates, and three to five ways of writing the question.
- The comparison. A group that went without the change separates its effect from the AI’s own growth, as in the glasp.co study.
- Who did it and what they sell. A study by someone who sells a tool can be good, if it says who paid.
To check an agency’s numbers, use the list in the guide what to ask before hiring a GEO agency.
murmur writes based on what the scientific studies support and measures the effect on each brand
murmur writes clients’ guides based on what the scientific studies support and measures the effect on each brand. From the 2023 article, murmur’s content production takes the cited sources, the quoted sentences from people who know the subject and the number with its origin in place of the vague description. From SAGEO Arena, it takes the answer in the first sentence and the buyer’s word in place of the technical term. From Martinez’s review, it takes the criterion for each number:
“The criterion is therefore not to “add numbers,” but to provide relevant, verifiable, dated, and properly attributed evidence.”
In other words, the number goes in when it is a true fact, that answers the question and brings the source and the date. murmur read the 2023 article, SAGEO Arena and Martinez’s review in full.
At murmur, a specialist dedicated to the brand studies the client’s sector and sees what the AI models answer the buyer, and who appears in place of the brand. They take care of the guides that answer that buyer’s questions. Week after week, the same questions go back to the AI models, and the dashboard included in the service shows the brand’s presence alongside that of competitors. Presence is the number of answers the brand appears in, divided by the total number of answers.
For clicks and sales, murmur’s answer has three parts. Most of the AI models’ influence arrives without a click, because few answers bring a source with a link. When there is a click, it shows up in the client’s own Google Analytics. murmur works on the part the click does not show: the brand’s presence in the answers, alongside each competitor, over time.
murmur’s case studies show that presence growing. Visio, operations software for retail, appeared in 2% of ChatGPT answers in June 2026 and reached 22% in July, with the same 135 questions in both months. In Claude, it went from 1% to 7%, and Google AIO answers with Visio went from 0% to 45%. Mindo, a São Paulo motion design studio, already had the entire guide published on 22 June 2026. In the June round, ChatGPT recommended it in 12% of the 134 questions in its market, and in the July round, in 23%. At FotoRegistro, an online photo printing store, the first part of the guide launched on 24 June 2026, and the first citation of it in ChatGPT came a month later, on 24 July. These and murmur’s other case studies are on the case studies page.
Frequently asked questions
Is a study’s percentage a target for my company?
No, because your company’s target comes from its own starting point and from the competitors in your market. First measure how many answers your brand appears in, with the step-by-step to measure brand presence in ChatGPT. Then compare each new measurement with that one, on the same questions.
Are the studies about GEO in AI models independent?
The two studies in this guide that went through review by other researchers had public support, and the ones done by companies deal with the company’s own site or system. Aggarwal’s had support from the National Science Foundation, the US government’s science agency, and SAGEO Arena, grants from the Korean government. Both authors of the glasp.co study work at Glasp, the company that owns the site, and the Pinterest study deals with the system Pinterest itself uses.
Is there a study in Portuguese about GEO and AI models?
There is one, which repeats the 2023 test in Portuguese. In August 2026, the founder of a Brazilian GEO agency published that preprint, with 200 questions on finance, health and travel, in a simulated search engine. What the studies in English are worth for pages in Portuguese is in the guide GEO works, for Portuguese content too.
The next study is your market’s
The studies say what increases a page’s space in the answer, and the answers the AI models give your buyer today say where to start. Just enter the site in murmur’s free diagnostic to see, without signing up and in under two minutes, the AI models’ answers to your market’s questions. To compare with other markets, murmur’s sector studies record a round of purchase questions and the brands that appeared in the answers.
Read also
- GEO works: what studies and case studies show about appearing in ChatGPT
- What is GEO: getting your company to appear in ChatGPT and AI answers
- How to write text to appear in ChatGPT: answer up front and numbers with sources
- Manipulating what ChatGPT recommends: it worked in a test and puts your site at risk
- Why AI recommends some companies and not others: the four reasons
References
- Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, “GEO: Generative Engine Optimization”, KDD 2024, arXiv 2311.09735, version 1 dated 16 November 2023 and version 3 dated 28 June 2024. Test on the one thousand questions in the test split of a 10 thousand-question dataset; the values are from Table 1, Overall column, and the gain by search position, from Table 2; the test on Perplexity used 200 examples, with the texts sent as a file. Support from the National Science Foundation, grant 2107048. https://arxiv.org/abs/2311.09735 (accessed on 3 October 2026).
- Sunghwan Kim, Wooseok Jeong, Serin Kim, Sangam Lee and Dongha Lee, “SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization”, KDD 2026, arXiv 2602.12187, version 1 dated 12 February 2026 and version 2 dated 7 August 2026. Yonsei and Konkuk universities, with grants from the Korean government (IITP). The structural fields, in Table 2. https://arxiv.org/abs/2602.12187 (accessed on 3 October 2026).
- Olivier Martinez, “Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)”, arXiv 2607.14035, 15 July 2026, preprint. Review of 45 studies from 16 November 2023 to 14 July 2026; the confidence of each claim is in Table 5, the support for each factor in Table 4 and the criterion for the numbers in section 7.2. https://arxiv.org/abs/2607.14035 (accessed on 3 October 2026).
- Keisuke Watanabe and Kazuki Nakayashiki, “Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic”, arXiv 2606.04362, 3 June 2026, preprint. The authors are from Glasp, the company behind the site studied; the set of changes is in section 4. https://arxiv.org/abs/2606.04362 (accessed on 3 October 2026).
- Faye Zhang, Qianyu Cheng, Jasmine Wan, Vishwakarma Singh, Jinfeng Rao and Kofi Boakye, “Generative Engine Optimization: A VLM and Agent Framework for Pinterest Acquisition Growth”, arXiv 2602.02961, 3 February 2026, preprint on the Pinterest system. https://arxiv.org/abs/2602.02961 (accessed on 3 October 2026).
- André Henrique de Paula, Portuguese-language replication of the 2023 GEO study, Zenodo, version 1.0, published on 10 August 2026, preprint without peer review. 200 queries on finance, health and tourism, in a simulated generative engine, with three repetitions per condition (6,000 runs). The author is the founder of Geostack, a Brazilian GEO agency. https://zenodo.org/records/21871669 (accessed on 3 October 2026).
- arXiv, search for the phrase “generative engine optimization” in the abstract, run by murmur on 3 October 2026: 50 results. https://arxiv.org/search/?query=%22generative+engine+optimization%22&searchtype=abstract
- murmur, free diagnostic: https://murmur.marketing/en/diagnostic/; sector studies: https://murmur.marketing/estudos/; case studies for Mindo, FotoRegistro, Fly Vet and Visio, published with the clients’ authorization: https://murmur.marketing/en/cases/, https://murmur.marketing/en/cases/mindo/, https://murmur.marketing/en/cases/fotoregistro/, https://murmur.marketing/en/cases/fly-vet/ and https://murmur.marketing/en/cases/visio/; guide on how to measure the brand’s presence in ChatGPT, with the number of answers that include a source: https://murmur.marketing/en/guide/how-to-measure-ai-brand-citations/ (accessed on 3 October 2026).