Updated on 7 November 2026, by João Leite, Founder of murmur.
Original research gives AI a number that only the company has. In the laboratory test that created the acronym GEO, adding statistics to a page increased its share of the AI answer, and in the Muck Rack count the AI models cited mostly earned media, such as the press. So publish the research with who measured it, how many people or questions, and when, and take the number outside the company’s website as well.
- An added number increased the page’s space in the answer. In the laboratory test that created the acronym GEO, the page’s share of the answer rose from 19.3 to 25.2 with added statistics, across a thousand questions.
- The number serves as proof when it is checked. In Martinez’s review of 45 studies, the number answers the question, has a date and says who measured it.
- Google asks, for its own search, whether the content brings original research. The OpenAI and Anthropic help pages for sites do not address the subject.
- In the Muck Rack count, the AI models cited mostly earned media, such as the press. That was 84% of the citations from ChatGPT, Claude and Gemini.
- Seven companies that publish their own data about AI answers in Brazil are part of the comparison: Brasil GEO, Escala Will, GEO Sounet, Geostack, murmur, naia and Ranqia.
murmur, a GEO agency in Brazil that places brands in the answers of ChatGPT, Claude, Google AIO and other AI agents, gives each client a dedicated specialist, who studies their market. murmur writes the pages that answer the buyer’s questions and tracks, week after week, what the AI models answer.
The comparison brings together seven companies that publish their own data about AI answers in Brazil, in alphabetical order. Each company comes with the link and the date of its page. murmur also publishes original research, in the sector studies, and is part of the comparison.
What is measured about original research and citation by AI
What is measured about original research is the effect of adding numbers to a page. The acronym GEO, for the work of making the brand appear in AI answers, came out of a 2023 laboratory test. In it, Pranjal Aggarwal and colleagues added statistics to a page, and its share of the answer rose from 19.3 to 25.2, across a thousand questions and with GPT-3.5.
The review of 45 studies that Olivier Martinez published in July 2026 replaces the idea of putting numbers in with a criterion. For Martinez, the number serves as proof when it answers the question, can be checked, has a date and says who measured it.
Surfer, which sells a text editor for SEO, compared the pages cited in more than 650 thousand AI answers. In ChatGPT, citing data and named sources was the factor most linked to citation in 2025, at 11%, and remained among the factors linked to it in 2026. Surfer itself warns that the calculation is a correlation and does not prove what causes the citation.
What Google, OpenAI and Anthropic say about original research and original content
Google asks whether the content brings original research; the three OpenAI and Anthropic help pages for sites do not address the subject. On Google’s page about helpful content, the first self-assessment question is whether the text brings original information, reporting, research or analysis. The fourth question asks whether someone writing from other sources avoids copying and adds value and originality.
In Google search, AI appears in features such as the AI Overview, text at the top of some results, and AI Mode, in which the person has a conversation. In Google’s guide to these features, first-hand evaluation brings a unique perspective, and a summary of what already exists only repeats. Both Google pages apply to its own search, and not to ChatGPT or Claude.
OpenAI, in its two help pages for publishers, asks that the site let OAI-SearchBot through, the ChatGPT search crawler. Anthropic’s page for site owners describes its three crawlers and what changes when the site blocks each one. None of the three pages, read on 3 October 2026, mentions original research or content type.
The types of original research a company does without an institute
Without an institute, a company does original research by asking people, reading its own data, comparing open data or repeating an index. Geostack lists these four formats on the same page from September 2026. They are the questionnaire with market respondents, the analysis of anonymized internal data, the analysis of open sector information, such as prices and response times, and the index in periodic editions. The seven companies in the comparison, murmur among them, ask the AI models what the buyer asks and count the brands in the answers.
How to publish original research so that AI can cite it
Publish the original research on an open page, in text, with each number in the sentence together with who measured it, the sample and the date. There are six precautions:
- The number with the sample, in the same sentence. Who measured it, how many people or questions, which AI models and the month appear alongside the number.
- The answer at the start of the page. The main finding goes in the first sentence, and the rest of the writing is in the guide on how to write a text to appear in ChatGPT.
- The table in HTML and the number in the text as well. Build the table in the page code, not as an image, and repeat the main numbers in a paragraph.
- The questions and the counting rule on the page. Publish the questionnaire or the list of questions and say what counted as a mention, so that anyone in doubt can redo the calculation.
- The same URL in every edition. In the check by Otterly, an AI monitoring company, 19.3% of the URLs cited by seven AI models in one month did not return a live page. Otterly itself writes that the calculation reflects the URLs checked up to that point, about one third of the set.
- The text ready in the HTML. An answer behind a tab, a button or JavaScript may not reach the AI crawler. The reason is in the part of the llms.txt guide about text ready in the HTML.
Original research should also leave the company’s website
Original research should also leave the company’s website and appear on sector sites and in news stories that repeat the number. Muck Rack, a public relations software company, separated more than 25 million links from the answers of ChatGPT, Claude and Gemini by type of source. In its count, from May 2026, 84% of citations went to earned media, and journalism alone accounts for 27% of the sources cited. Earned media is what others publish, such as the press, and the page does not define the category. In the same count, paid content or advertorial got 0.3% of citations. Among media sites, the sector specialist comes before the general press: in the count by Otterly, a company that sells AI monitoring, made with Press Ranger and updated in August 2026, the niche or sector outlet took 75.8% of media citations, against 24.2% for the large outlet, across 129.3 million citations of questions from the United States. And in August 2026 ChatGPT halved the share of journalism in its citations, while the other engines varied within 7%, in Otterly’s AI Search Index, updated in September 2026, with 6,095,520 citations from 17 sectors of the United States.
How to compare original research about AI answers
Compare original research about AI answers by the method written on the page, with the same eight criteria for all:
- who measured, and whether they sell what they measured;
- what was counted: mention, position, recommendation or citation as a source;
- how many questions, and whether they include the name of any brand;
- which AI models, each with its own number;
- the collection dates and how many rounds;
- how the questions were asked: on screen, as the buyer sees it, or via API, when a program asks the AI directly;
- what the reader can check: the questions, the stored answers and the data;
- what the author themselves writes that the study does not show.
Each company counts in its own unit, such as answer, query, test or observation, and a number can only be compared with another number in the same unit.
Who publishes their own data on AI answers in Brazil
Brasil GEO, Escala Will, GEO Sounet, Geostack, murmur, naia and Ranqia publish their own data on AI answers in Brazil. Besides murmur itself, these are the six companies murmur found up to 3 October 2026, on the web and on the study pages of GEO companies.
Brasil GEO, a GEO consultancy from Goiânia, keeps a historical dashboard of AI answers about brands. On the home page, reviewed on 5 October 2026, the September snapshot brings together 92,064 AI answers about 127 entities, with the last collection on 11 September. Among the entities, 16 are fictitious brands, which serve as a control. The 92,064 answers come from six AI models: ChatGPT, Claude, Gemini, Perplexity, Grok and Llama 3.3, retired in August.
Brasil GEO details the measurement of mentions in BRGEO-1, a preprint of 11 September 2026, a text deposited before review by other researchers. Its inventory brings together 91,392 answers obtained via API, on 56 dates between 23 April and 11 September 2026. The home page notes that the inventory and the dashboard have different counts. The abstract says that the study does not establish a universal ranking of models and that the complete set of data per answer was not published in this version. The author, Alexandre Caramaschi, is affiliated with Brasil GEO, which provides services on the topic.
Escala Will is a score from 0 to 100 for brand visibility in AI models, created by Wilson Silva at ESPM São Paulo. The author’s page brings 552 brands from 22 sectors and 6,611 queries to ChatGPT, Gemini, Perplexity and Claude, with numbers published in August 2026. The collection took about six months, and the questions never named the brand directly.
The Escala Will page says that the protocol, the data and the analysis plan have been on the OSF, an open research repository, since August 2026, without giving the URL of the registration. Wilson Silva is CEO of WS Labs, which owns Apareci.ai, subscription-based software that calculates the same score.
GEO Sounet, which sells GEO services, published the Estado do GEO no Brasil 2026 on 22 July 2026. The page describes 200 companies from 10 segments, 5 standardized questions (prompts) per segment and 4 AI models, ChatGPT, Gemini, Copilot and Claude: 4,000 tests. The study records citation, position and context. The page does not include the questions, the collection date, the way the answers were obtained or the limitations of the study.
Geostack keeps online the Citation Share Report on digital banks, of 29 April 2026, updated on 27 September. The report poses 200 questions to five platforms, ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews, once each, for a total of 1,000 observations. In internal tests, Geostack repeated 30 questions five times on each platform and measured an average stability of 84% of brand mentions between rounds, in financial verticals. The counting rule for each stage of the purchase is written down, and some of the example questions include the name of a bank. The collection date does not appear, and Geostack writes that the findings reflect the state of the AI models at the time of the study.
Geostack also publishes a Portuguese-language replication of the 2023 test that created the acronym GEO. The preprint is by André Henrique de Paula, the agency’s founder, and has been on Zenodo, an open research repository, since 10 August 2026. The test runs on an AI search simulated by the author, and the Geostack page says that the 6,000 runs are published alongside the article. The Zenodo record lists only the PDF.
In the September 2026 sector studies, murmur asked 25 purchase questions, without company names, to three or four AI models. Each question was asked once, in a temporary chat with no history, from an IP in Brazil. Only the name the AI writes in the answer counts. In the demi-fine jewelry study, there were 75 answers from ChatGPT, Claude and Google AIO, all saved as screenshots. The page publishes the count for each brand that appeared, the 25 questions in full and what the study does not show. The demi-fine jewelry study page says that no company commissioned, reviewed or paid for it, and murmur also sells GEO services. Each September 2026 study records a single round.
naia, which sells a platform for monitoring brands in AI models, publishes the naia Index, a monthly list of the brands most cited by category. According to the methodology updated on 4 October 2026, the same five questions go to ChatGPT and Gemini every month, with web search turned on. Each category totals 10 queries per round, with the instruction to list up to 10 brands. The page for each category shows the questions sent, there is a public API, and the methodology says that each number can be reproduced from the stored original answer. naia notes that the index does not measure product quality and is not market share.
Ranqia, in early access, offers to measure what AI models answer and to place brand content where the AI looks. In the Ranqia Papers, it lists 18 publications, from 25 June to 24 September 2026, all about ChatGPT. In the card machine study, from 20 August 2026, there are 39,687 ChatGPT answers to 50 questions, from 772 to 800 per question, in the August window. The page shows a sample of the questions and results, and the method notes are in the PDF. Ranqia writes that the indicator measures frequency and order of appearance, and not the actual rate, settlement period, contractual conditions or commercial recommendation.
| who publishes | what the page reports | brand names in the questions | AI models | when | screen or API | what the reader can check on the page |
|---|---|---|---|---|---|---|
| Brasil GEO | 92,064 AI answers about 127 entities, 16 of them fictional, in the September 2026 cut | the page does not say | ChatGPT, Claude, Gemini, Perplexity, Grok and Llama 3.3 (retired on 19 August 2026) | last collection on 11 September 2026; preprint inventory from 23 April to 11 September 2026 | API, in the preprint | the method, in the preprint abstract; the full per-answer dataset was not published in this version |
| Escala Will | 6,611 queries about 552 brands from 22 sectors | no, according to the page | ChatGPT, Gemini, Perplexity and Claude | about six months; OSF registration in August 2026 | the page does not say | the OSF registration address: the page does not say |
| GEO Sounet | 4,000 tests: 200 companies, 5 questions per segment, 4 AI models | the page does not say | ChatGPT, Gemini, Copilot and Claude | published on 22 July 2026; collection date: the page does not say | the page does not say | a summary of the method; the questions: the page does not say |
| Geostack | 1,000 observations: 200 questions across 5 platforms, once each | yes, in part of the example questions | ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews | published on 29 April 2026 and updated on 27 September 2026; collection date: the page does not say | the page does not say | the counting rule for each step and example questions |
| murmur | 75 answers: 25 questions, once in each AI model, in the demi-fine jewelry study | no | ChatGPT, Claude and Google AIO | September 2026, one round | screen, in a temporary chat with no history | the 25 questions in full, the count of each brand that appeared and what the study does not show |
| naia | 10 queries per category in each round: 5 questions across 2 AI models | no, in the questions the page shows | ChatGPT and Gemini, with web search | every month, with a reference date on day 1; methodology of 4 October 2026 | the page does not say | the questions sent, the methodology and a public API |
| Ranqia | 39,687 answers to 50 questions, in the card machine study | the page does not say | ChatGPT | August 2026 window; 18 publications from 25 June to 24 September 2026 | the page does not say | a sample of the questions and results, and the PDF |
The seven companies differ on four method points. murmur and naia show the questions, murmur in full and naia the ones for each category, and Geostack and Ranqia show examples or a sample. The Brasil GEO, Escala Will and GEO Sounet pages do not include the list. Only Escala Will, Geostack and murmur state whether the questions include a brand name.
Only Brasil GEO says, in the preprint, that the answers came from an API, and only murmur says they came from the screen. The pages of the other five do not say whether the collection was done via the screen or via API. Brasil GEO, Geostack, murmur, naia and Ranqia state what the study does not show, and the Escala Will and GEO Sounet pages do not. All seven sell a GEO service or tool, or are linked to someone who does.
murmur writes the client guides and publishes the questions of each study
murmur writes the guides that answer each client’s buyer and publishes the questions of each sector study. The guides live at a URL on the client’s brand domain, a subdomain, and the content belongs entirely to the client, even after the contract ends. Four clients have a published case study with their name: Mindo, Visio, FotoRegistro and Fly Vet. ChatGPT began recommending Mindo, a motion design studio, in 23% of 134 questions in July 2026, against 12% in June. On Google AIO, the AI Overview, Visio went from 0% to 45% of answers on the same 135 questions, from June to July 2026.
In the client service, the same set of questions goes back to the AI models week after week. The included dashboard shows the brand’s presence alongside that of competitors. Presence is how many answers the brand appears in, divided by the total number of answers.
Frequently asked questions
Does AI read PDFs and e-books?
Google indexes PDF and ePub; the three pages from OpenAI and Anthropic for sites do not say whether their crawlers read these files. PDF and ePub are on the list of files Google indexes. Put the numbers on a page, in text, and leave the PDF for the full study. The industrial catalog in PDF, and what to do with the technical sheet that exists only as an image, is in catalog in PDF and technical sheet as an image.
Do I need to redo the research from time to time?
Your own research needs to be redone, because AI answers change over time. naia redoes the list every month, using the methodology of 4 October 2026. In the report updated on 27 September, Geostack recommends rereading the financial sector at least every three months. murmur’s sector studies record one round. In the client service, murmur asks the AI models the same questions week after week.
Can you do your own research without a research institute?
Your own research depends on the question, the written base and the data you keep, not on an institute. An institute is only needed when the number has to speak for an entire audience, with a random sample.
Can I trust a study by someone who sells what they measured?
A study by someone who sells what they measured is worth it when the method is on the page. The method includes the questions, the AI models, the dates, the counting rule and what the study does not show. To choose a supplier, the list of what to ask is in the guide on what to ask before hiring a GEO agency.
Where to start your own research
Start with the question your buyer asks that still has no published number, and write the base before collecting. To see a method already published, with the questions, the AI models, the month and what the study does not show, open murmur’s demi-fine jewelry study. murmur’s free diagnostic starts from your site’s URL and returns, with no sign-up, the AI answers in your market.
Read also
- How to write text to appear in ChatGPT: answer up front and numbers with sources
- How to get your company cited by AI models: ChatGPT, Claude and Google AIO
- Scientific studies on GEO: the effect on the AI answer
- Brand presence in AI models: the reliable metric and how to check any number
- How many Brazilian companies appear in AI answers: it depends on the sector and the AI model
References
- Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, “GEO: Generative Engine Optimization”, KDD 2024, arXiv 2311.09735. The values are from Table 1, Overall column, with GPT-3.5, on the one thousand questions in the test portion of a bank of 10,000. https://arxiv.org/abs/2311.09735 (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. A review of 45 studies; the criterion for the evidence is in section 7.2. https://arxiv.org/abs/2607.14035 (accessed on 3 October 2026).
- Surfer, Paulina Kaleta, “How to Rank in ChatGPT: What the Data Actually Says”, updated on 17 September 2026. More than 650 thousand answers from real interfaces, in collections from October 2025 and April 2026; correlation data, in the words of Surfer itself, which sells a text editor for SEO. https://surferseo.com/blog/how-to-rank-in-chatgpt/ (accessed on 3 October 2026).
- Google, “Como criar conteúdo útil, confiável e que prioriza as pessoas”, Google Search Central, Portuguese version last updated on 18 December 2025, section “Perguntas sobre o conteúdo e a qualidade”. https://developers.google.com/search/docs/fundamentals/creating-helpful-content?hl=pt-br (accessed on 3 October 2026).
- Google, “Otimizar seu site para recursos de IA generativa na Pesquisa Google”, Google Search Central, Portuguese version last updated on 14 July 2026, section “Crie conteúdo valioso e exclusivo para seu público”. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide?hl=pt-br (accessed on 3 October 2026).
- Google, “Tipos de arquivos indexáveis pelo Google”, Google Search Central, Portuguese version last updated on 6 February 2026. https://developers.google.com/search/docs/crawling-indexing/indexable-file-types?hl=pt-br (accessed on 3 October 2026).
- OpenAI, Help Center, “Searching the web with ChatGPT” and “Publishers and Developers - FAQ”, undated page. https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt and https://help.openai.com/en/articles/12627856-publishers-and-developers-faq (read in the browser on 3 October 2026).
- Anthropic, “Does Anthropic crawl data from the web, and how can site owners block the crawler?”, Claude Help Center, updated on 7 April 2026. The page describes ClaudeBot, Claude-User and Claude-SearchBot. https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler (accessed on 3 October 2026).
- Muck Rack, “Earned media still drives 84% of AI citations. Here’s what that means for PR.”, 7 May 2026. More than 25 million links cited by ChatGPT, Claude and Gemini, across 17 sectors; the earned media classification is Muck Rack’s, which sells public relations software, and the page does not define the category. https://muckrack.com/blog/what-is-ai-reading-may-2026 (accessed on 3 October 2026).
- Otterly with Press Ranger, “News Publishers With OpenAI Deals Get +48% More AI Citations on ChatGPT”, updated on 21 August 2026. 129.3 million citations from June 2026, across seven AI models, with questions from 16 sectors in the United States; the split of 75.8% against 24.2% is between niche or sector outlets and large outlets, within news and media citations. Otterly sells AI monitoring. https://otterly.ai/blog/ai-news-licensing-deals/ (accessed on 6 October 2026).
- Otterly, Rick Tousseyn, “The AI Search Index: What ChatGPT, Gemini, Perplexity and Google Actually Cite in 17 Industries”, updated on 8 September 2026. 6,095,520 citations from July and August 2026, across seven AI models, with questions from 17 sectors in the United States; the news and media share in ChatGPT fell in two steps, on 8 and on 14 August. Otterly sells AI monitoring. https://otterly.ai/blog/ai-search-index/ (accessed on 7 October 2026).
- Otterly, Pam Tarrayo, “Dead Citations: 19.3% of AI Search Sources Are No Longer There”, 24 July 2026, modified on 14 September 2026. The table adds up to 20,092,769 addresses cited in one month by seven AI models, and the scope note says the counts reflect the addresses checked up to that point, about one third of the set; “not live” includes unreachable, redirect and server error; Otterly sells AI monitoring. https://otterly.ai/blog/dead-citations-study/ (accessed on 5 October 2026).
- Geostack, André HP, “Pesquisa Proprietária: Como Dados Originais Transformam Sua Marca em Fonte Primária para LLMs”, 7 September 2026, updated on 27 September 2026. https://geostack.com.br/pesquisa-proprietaria-como-dados-originais-transformam-sua-marca-em-fonte-primaria-para-llms/ (accessed on 3 October 2026).
- Geostack, André HP, “Citation Share Report, Bancos Digitais no Brasil: como ChatGPT, Perplexity, Gemini, Claude e Google AI Overviews recomendam fintechs e contas digitais”, 29 April 2026, updated on 27 September 2026. The recommended interval is in question 14 on the page: “Trimestral, no mínimo” (Quarterly, at minimum). https://geostack.com.br/citation-share-report-bancos-digitais-no-brasil-como-chatgpt-perplexity-gemini-claude-e-google-ai-overviews-recomendam-fintechs-e-contas-digitais/ (accessed on 5 October 2026).
- Geostack, “Pesquisa GEO no Brasil: Estudo Científico Publicado por André HP”, 10 August 2026, updated on 31 August 2026. https://geostack.com.br/pesquisa-geo-no-brasil-estudo-cientifico-publicado-por-andre-hp/ (accessed on 3 October 2026).
- André Henrique de Paula, “Otimização para Motores Generativos em Português Brasileiro: uma replicação do estudo GEO de Princeton em consultas de finanças, saúde e turismo”, Zenodo, version 1.0, 10 August 2026, preprint. 200 queries and 6,000 runs in a simulated search; the record lists one file, artigo_geo_brasil.pdf, of 271.1 kB. https://zenodo.org/records/21871669 (accessed on 3 October 2026).
- Brasil GEO, “Brasil GEO | Consultoria de GEO e pesquisa em IA”, revised on 5 October 2026, with the excerpt from the September 2026 historical dashboard and the last collection on 11 September 2026. https://brasilgeo.ai/ (accessed on 5 October 2026).
- Brasil GEO, “BRGEO-1: como medir menções em respostas de IA”, summary of 18 September 2026 of the preprint by Alexandre Caramaschi deposited in Zenodo on 11 September 2026, DOI 10.5281/zenodo.22711743. https://brasilgeo.ai/pesquisa/protocolo-brgeo-1 (accessed on 5 October 2026).
- Wilson Silva, page of Escala Will, with the numbers attributed to E-Commerce Update of 17 August 2026. https://wilsonsilva.net.br/escala-will.html (accessed on 3 October 2026).
- WS Labs, page of Wilson Silva’s company, which lists Apareci.ai among its products as a subscription software for GEO and states that the founder’s academic affiliation does not represent ESPM’s endorsement of the products. https://wslabs.ai/ (accessed on 3 October 2026).
- GEO Sounet, “Estado do GEO no Brasil 2026: pesquisa original”, 22 July 2026. https://geo.sounet.com.br/blog/estado-do-geo-no-brasil-2026/ (accessed on 3 October 2026).
- naia, “naia Index: marcas mais citadas por IA” and “Como o naia Index calcula o ranking público”, methodology in version 5, updated on 4 October 2026; on 5 October 2026, the index page showed 55 active categories. https://naia.today/naia-index and https://naia.today/naia-index/metodologia (accessed on 5 October 2026).
- naia, “Planos naia: Autoridade e Enterprise para visibilidade em IA”, pricing page. https://naia.today/pricing (accessed on 3 October 2026).
- Ranqia, “Estudos originais sobre visibilidade em IA”, the two pages of the Ranqia Papers, and “Quais maquininhas o ChatGPT recomenda no Brasil?”, 20 August 2026. https://www.ranqia.ai/papers/ and https://www.ranqia.ai/papers/maquininhas-que-o-chatgpt-recomenda/ (accessed on 3 October 2026).
- murmur, study “O que as IAs respondem quando alguém procura uma marca de semijoias”, September 2026, and the index of sector studies: https://murmur.marketing/estudos/semijoias-2026-09/ and https://murmur.marketing/estudos/; case studies for Mindo, Visio, FotoRegistro and Fly Vet, published with the clients’ authorization: https://murmur.marketing/en/cases/mindo/, https://murmur.marketing/en/cases/visio/, https://murmur.marketing/en/cases/fotoregistro/ and https://murmur.marketing/en/cases/fly-vet/; free diagnostic: https://murmur.marketing/en/diagnostic/ (accessed on 3 October 2026).