Updated on 19 October 2026, by João Leite, Founder of murmur.
To appear in ChatGPT, write each page about one buyer question, answer it in the first sentence and keep the numbers in the text, with the source. The title and the headings use the words of whoever is asking. Price, timeframe, name and date stay written in the body of the page, outside images and buttons. The published tests, run on ChatGPT and in the lab, point the same way: the right subject, with proof that can be checked.
- The answer comes in the first sentence. In Surfer’s measurement, the three factors most linked to citation in ChatGPT look at the start of the text.
- The most common form of the question opens the page. It goes in the title and in the first sentence, and the other forms go in the body.
- The number goes in the text, with who measured it. In a study by Zhang Kai and colleagues, pages with numbers contributed, on average, 61.55% more to the answer, in a comparison between pages.
- Data outside the text disappears. In an Otterly test, only 3 of 120 answers found a piece of data placed in the file name and in the image description. None found what was written inside it.
- Comparison and step by step appear in the answer more often. In the same study by Zhang Kai and colleagues, pages with a comparison contributed, on average, 55.28% more to the answer, and those with a step by step, 41.20%.
To appear in ChatGPT, the answer comes in the first sentence
The first sentence after the title answers the buyer’s question, with the subject named and no introduction. The test is to cut out that sentence and read it alone: if it does not answer, it is still a warm-up. The rest of the first paragraph completes the answer, and anyone who reads only that comes away with the essentials.
Surfer, a company that sells a text editor for SEO, scored the sources of more than 650 thousand AI answers against a list of 20 factors. In ChatGPT, the three factors most linked to citation in 2026 look at the start of the text: whether it confirms the question, whether it serves the reason behind it and whether it gives the short answer before the rest. Surfer notes that this is correlation data.
In the lab, the position of the answer also matters. In a case study from SAGEO Arena, by Kim and others, accepted at KDD 2026, a data science conference, the page with the answer right at the start moved up in the reranking. Reranking is the stage in which the search chooses the pages the AI will read.
Each paragraph also needs to make sense on its own. In the Bing guidelines for webmasters, written by Microsoft, the owner of this search engine that OpenAI cites among ChatGPT’s search partners, it says that a page tends to become a source more often when the text makes sense by itself, brings the essentials up top and does not mix unrelated topics. Anthropic, the company that makes Claude, gives the example of the sentence “the company’s revenue grew 3% compared with the previous quarter”, which, cut out, does not say which company or which quarter. Writers avoid the problem by opening each paragraph with the name of the subject, and not with “this” or “it”.
murmur, a GEO agency in Brazil that places brands in the answers of ChatGPT, Claude, Google AIO (the summary written by AI that Google shows at the top of many searches) and other AI agents, writes the FotoRegistro guide. The store sells photo printing over the internet. The guide page about printing photos online opens with the whole answer, in 23 words. The sentence is this: “Vale a pena imprimir fotos online quando o site imprime em fábrica própria, mostra o mockup antes do pagamento e tem reputação verificável.” (“Printing photos online is worth it when the site prints in its own factory, shows the mockup before payment and has a verifiable reputation.”) Pulled out of the page, it keeps answering.
The most common form of the question opens the page, and the others go in the body
Each page uses the most common form of the question in the title and in the first sentence, and the other forms stay in the body. Pages are divided by the questions that actually differ for the buyer, each one answered in full, and wording variations of the same question do not get their own URL. A page for one question can only answer it in the first sentence; the one that covers everything has nowhere to start.
On the page that answers “como tirar o visto americano” (“how to get a US visa”), for example, “visto dos Estados Unidos” (“United States visa”) and “visto para os EUA” (“visa for the USA”) appear in the body and in a frequently asked question. How to tell whether two questions need different pages, and how many to publish, is in the guide on how many pages to publish for your company to appear in ChatGPT.
Title and headings with the buyer’s words, and each number with who measured it
The title and the headings repeat the words of the buyer’s question, and each number goes into the text with who measured it. The title is the H1, the main heading of the page; sections carry H2, and subdivisions H3, marked in the HTML, the code that forms the page. Each H2 states the answer to a doubt that comes after the main question. The URL has the words of the question, without a date, without a number and without a parameter, the part after the ”?” that some sites add.
Among the pages ChatGPT finds for a question, it cites more of those whose title is similar to it. Ahrefs, which sells SEO tools, measured this across 1.4 million questions asked of ChatGPT, with open models that approximate its calculation. On an index that goes up to 1, the similarity between title and question was 0.602 on cited pages and 0.484 on those left aside. URLs written with words were cited 89.78% of the time, against 81.11% for URLs without them. In a measurement by Otterly, a company that sells AI monitoring, across 1,028,959 URLs cited by six AI models in May 2026, the URL without a parameter had 2.1 citations on average, against 1.6 with a parameter. The whole path of source selection is in the guide on how ChatGPT search works.
The buyer’s words also count in the body of the text. SAGEO Arena tested 171,003 real pages in a search built by the authors. There, the first selection matches words, and the page with the common word, such as “eating”, receives a higher relevance score than the page with the technical term, such as “alimentary routines”. The authors explain that the question uses the common word.
Each number goes into the running text, the ordinary paragraph, with who measured it, over how many and when, in the same sentence or the next. That way it can be checked and stays right when the AI cuts it out. “Ahrefs measured 1.4 million questions asked of ChatGPT” says who and how many; “most pages” does not.
Zhang Kai and two colleagues measured, in an April 2026 study, how much each cited page enters the final answer. There were 18,151 pages cited by ChatGPT, by Google’s AI Overview and by Perplexity. Pages with numbers or statistics contributed, on average, 61.55% more to the answers than pages without them. It is a comparison between pages, across the three platforms combined, and not a test of cause.
In the study by Aggarwal and others, which created the term GEO and was presented at KDD 2024, a page rewritten with statistics went from 19.3% to 25.2% of the answer, on a measure that discounts pages appearing further down. With sentences in quotes from sources, it went to 27.2%, and with the sources cited, to 24.6%. The test used a thousand questions and GPT-3.5, an OpenAI model, reading the first five Google results. The review of 45 studies by Olivier Martinez, from July 2026, sums up the criterion:
“The criterion is therefore not to “add numbers,” but to provide relevant, verifiable, dated, and properly attributed evidence.”
For Martinez, what counts is not putting numbers in, but proof tied to the question, which can be checked, with a date and with the author named. He recalls that an invented number may even be used more by the AI and, even so, leave the answer worse. Anyone who wants to publish their own research will find the step-by-step in the guide on original research to appear in ChatGPT.
Table, list and summary organize the text, and comparison and step-by-step appear more often in the answer
Table, list and summary organize the data for the reader, and the page that compares options or lays out the steps in order appears more often in the answer. In the study by Zhang Kai and colleagues, pages with a comparison contributed 55.28% more to the answers, on average, and those with a step-by-step, 41.20%, across the three platforms combined.
In a test by Vishwakarma and colleagues, presented at SIGIR 2026, an information retrieval conference, six AI models chose between two sources in 252,000 rounds, and the sources differed on a single point. The six models decided based on the subject, the price shown, the 2026 date versus the 2019 date and the order of the sources. The test is a laboratory one, with two product reviews pasted into the conversation, not ChatGPT’s search.
The table works when the reader compares items row by row. It is built in HTML, not as an image, and its conclusion also goes in a paragraph. The bullet-point summary comes right after the first paragraph, with numbers that are also in the body.
The length of the text comes from the question, and the very short page is the least used
The right length is what the question asks for, and, among the cited pages, the shortest ones were the least used in the answer. In the study by Zhang Kai and colleagues, the most used quarter of pages had, on average, 1,943 words, and the least used quarter, about 170. The authors warn that long text does not win on its own: length helps together with useful structure, adherence to the question and proof.
Schema and questions at the end of the page: ChatGPT uses the page text, not the markup
ChatGPT uses the page text, and, in the published tests, data that was only in the markup was not found. Structured data, or schema, is markup in the page code that tells machines what each piece of information is, such as name, price and address. JSON-LD is a way of writing that markup, in a block the visitor does not see, and FAQPage is the markup for frequently asked questions.
In a test by searchVIU, an SEO tools company, five AI models were consulted in October 2025 about the eight prices on a page, each in a different format, reading it directly. The price that was only in the JSON-LD was not found by any of them, and ChatGPT found the three that were visible. In the Otterly test, which sells AI monitoring, an answer that existed only in the frequently asked questions markup on its site was not used by any of the seven AI models tested. Anyone who maintains the markup puts in it the same text that is on the screen.
The questions at the end of the page are worth it because of the answer written right below, with a fact the page had not yet stated. Write four to six questions the way the buyer asks, with the answer in visible text, ready in the HTML the server delivers. Why text assembled afterwards by JavaScript, in the browser, may not reach the AI is in the guide about llms.txt.
What matters most for a text to be cited, according to the studies
To be cited, the text answers the person’s question with proof that can be checked. OpenAI asks publishers to let OAI-SearchBot, the crawler that reads sites for ChatGPT’s search, into the site.
Olivier Martinez’s review of 45 studies gives each lever a degree of support. Answering the question has strong support in controlled tests, and proof that is easy to extract, such as a number, a definition and a comparison, has moderate to strong support.
| study | how it was done | what it asks of the page |
|---|---|---|
| Aggarwal and others, KDD 2024 | GPT-3.5 via API, reading the first five Google results, on a thousand questions | sentences in quotes, statistics and cited sources |
| SAGEO Arena, KDD 2026 | search built by the authors, over 171,003 real pages | the buyer’s words in the body, in the title and in the headings, and the answer at the beginning |
| Zhang Kai and colleagues, 2026 | 18,151 pages cited by ChatGPT, by AI Overview and by Perplexity | number, definition, comparison and step by step inside the text |
| SIGIR 2026, 252,000 rounds | six models via API, two sources at a time | staying on the subject, writing the price and keeping the date current |
| Martinez, July 2026 | review of 45 studies | a relevant, complete page that can be checked, organized and legible to search |
Read together, the five studies in the table ask for the same page: the answer at the beginning, with the words of the person asking, and numbers, definitions and comparisons in the text, with the source. Format and markup help organize, and none of them replaces the proof. How much each factor matters for the chance of a page becoming a source is in the guide about what AI citability is. The tests, one by one, are in the guide about scientific studies on GEO. The verdict is in the guide GEO works: what studies and case studies show about appearing in ChatGPT. If your company does not appear in any answer, first check whether the problem is the text, in the guide about why ChatGPT does not recommend your company.
Text on your site and mentions on other sites do different jobs
The text on your site and the mention on other sites do different jobs, and one does not replace the other. Your own page is the one the AI can cite as an answer, with your words and your numbers. What other sites write carries weight when the AI recommends a company, and the studies on that weight are in the guide about why the AI recommends some companies and not others.
The mention of the brand name, with or without a link, goes along with presence in AI answers. Presence is how many answers the brand appears in, divided by the total number of answers. In an Ahrefs study with 75,000 brands, from December 2025, the number of links to the site barely went along with the brand appearing in AI answers. Mentions of the brand name went along much more closely. It is correlation, and Ahrefs warns that it does not prove cause. The guide about backlinks for appearing in AI shows what was measured, in each AI, about links from other sites to yours.
At murmur, the sector study decides what to write, and the guides answer the questions of the client’s sector
At murmur, the client’s sector study decides what is written, and the guides answer the questions of their sector. In the Mindo guide, a motion design studio in São Paulo, the page about video for corporate training opens with the definition and includes a bullet-point summary right below. The writing follows academic studies in the field, such as Aggarwal’s, SAGEO Arena and Martinez’s review. The guide sits on a subdomain, a URL within the brand’s domain, such as conteudo.fotoregistro.com.br, and the content remains the client’s if the contract ends.
Frequently asked questions
Can you write this text alone, with the help of an AI?
The best option is to write with a GEO agency such as murmur. Before writing, murmur turns to data engineering to gather in-depth information about the company and the comparison with competitors, and then measures reliably and delivers results. Writing alone with an AI at your side, without data and without measuring, can be useful, but it rarely leads to the same result. The work requires a routine that does not stop: finding the buyer’s questions, checking the source of every number and updating the page. After that, you need to measure regularly whether the company started appearing in ChatGPT.
At murmur, a specialist dedicated to the brand studies the sector, takes care of the pages and tracks which AI answers the brand appears in.
Does the AI read the text inside an image?
Do not count on it: in an Otterly test, no AI model extracted a piece of data written only inside the image. Otterly, which sells AI monitoring, put an invented piece of data only in the image and in the texts linked to it, in six versions of a page, and ran 120 rounds in five AI models. Only 3 answers got it right, all in ChatGPT, with the data in the file name and the alt text together. Alt text, or alt, is the description of the image that sits in the code. More than half of the answers invented a number, and that is why the data also goes in the text.
Do who signs the text and the author page matter?
The byline and the author page matter so that readers know who is accountable for the text. Sign with name, role and date, right below the title. The about page repeats the same name and the same description as everywhere else, which helps the AI recognize the company without confusing it with another.
Does changing the date on the text make the AI cite it more?
On its own, a new date is not enough: what counts is a real update, visible on the page. In the lab test by Vishwakarma and colleagues, the page dated 2026 beat the one dated 2019 across all six models, with everything else equal. That is why the date has to be true. In a July 2025 study, Ahrefs looked at 17 million citations, from AI models and from regular Google search. The pages ChatGPT cited were, on average, 458 days newer than those in regular Google search, and the author warns that weak text updated every day does not gain ground by magic.
In the Ahrefs study with 1.4 million questions to ChatGPT, the conclusion was that a new date alone is not enough, and that relevance is what decides between the pages found. For its search, Google asks for a visible date, with Publicado (“Published”) or Atualizado em (“Updated on”), matching the date in the markup. Changing only the date is not an update, and what is said about that trick is in the guide on manipulating ChatGPT.
Where to start writing
Start with the question your buyer asks most and the page that should answer it. If it does not exist, or does not answer in the first sentence, it is the first text to write. Before writing, check in murmur’s free diagnostic what the AI models answer when asked the questions in your market. With no sign-up and in under two minutes, it shows the answers as they come in. murmur’s sector studies include, for several markets, a round of purchase questions with the brands the AI models named in it.
Read also
- How many pages to publish so your company shows up in ChatGPT: one per unanswered question
- llms.txt: What It Is, What It Is For, and Whether It Works to Get Cited by AI
- How ChatGPT search works: from the question to the sources in the answer
- Why ChatGPT does not recommend your company, and what to show the boss
- How to get your company to appear in ChatGPT: the step by step, with no registration or ads
References
- Surfer, Paulina Kaleta, “How to Rank in ChatGPT: What the Data Actually Says”, updated on 17 September 2026. More than 650 thousand answers, in rounds from October 2025 and April 2026; correlation data, in Surfer’s own words. https://surferseo.com/blog/how-to-rank-in-chatgpt/ (accessed on 1 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, 12 February 2026, revised on 7 August 2026. The sentence about the answer at the beginning comes from the case study in Figure 3. https://arxiv.org/abs/2602.12187 (accessed on 1 October 2026).
- Microsoft, “Bing Webmaster Guidelines”, Bing Webmaster Tools documentation, undated page. It applies to Bing and Copilot, and Bing is ChatGPT’s search partner. https://www.bing.com/webmasters/help/webmaster-guidelines-30fba23a (accessed on 1 October 2026).
- Anthropic, “Introducing Contextual Retrieval”, 19 September 2024. https://www.anthropic.com/engineering/contextual-retrieval (accessed on 1 October 2026).
- FotoRegistro, guide written by murmur, “Vale a pena imprimir fotos online? Qualidade, segurança e prazo”, updated on 8 July 2026. https://conteudo.fotoregistro.com.br/vale-a-pena-imprimir-fotos-online/ (accessed on 1 October 2026).
- Ahrefs, Louise Linehan, “Why ChatGPT Cites One Page Over Another (Study of 1.4M Prompts)”, 15 April 2026. The similarity was calculated with open models, as an approximation of ChatGPT. Ahrefs sells SEO and AI monitoring tools. https://ahrefs.com/blog/why-chatgpt-cites-pages/ (accessed on 1 October 2026).
- Otterly, “AI Citation Study: What URL Structures Get Cited Most (1M URLs)”, 7 May 2026. 1,028,959 URLs cited by six AI models. Otterly sells AI monitoring. https://otterly.ai/blog/url-ai-citations-study/ (accessed on 17 September 2026).
- Zhang Kai, He Xinyue and Yao Jingang, “From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms”, arXiv 2604.25707, 28 April 2026, version 2 of 29 April 2026, preprint. 602 questions, 21,143 citations and 18,151 pages downloaded, with the public data from geo-citation-lab. https://arxiv.org/abs/2604.25707 (accessed on 1 October 2026).
- Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, “GEO: Generative Engine Optimization”, KDD 2024, arXiv 2311.09735, version 1 of 16 November 2023 and version 3 of 28 June 2024. The values are from Table 1, Overall column of the position-adjusted word count, in the test portion of one thousand questions from a bank of 10 thousand. https://arxiv.org/abs/2311.09735 (accessed on 1 October 2026).
- Olivier Martinez, “Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)”, arXiv 2607.14035, 15 July 2026. The level of support for each lever is in Table 4. https://arxiv.org/abs/2607.14035 (accessed on 1 October 2026).
- Rahul Vishwakarma, Shushant Kumar and Ratnesh Jamidar, “What Gets Cited: Competitive GEO in AI Answer Engines”, SIGIR 2026, arXiv 2605.25517, 25 May 2026. https://arxiv.org/abs/2605.25517 (accessed on 1 October 2026).
- searchVIU, “Schema Markup and AI in 2025: What ChatGPT, Claude, Perplexity & Gemini Really See”, 2 December 2025. Test of 30 October 2025 on a page of searchVIU’s own, which sells SEO tools. https://www.searchviu.com/en/schema-markup-and-ai-in-2025-what-chatgpt-claude-perplexity-gemini-really-see/ (accessed on 1 October 2026).
- Otterly, Rick Tousseyn, “GEO Experiment: Does Schema Markup Really Impact AI Search?”, updated on 23 March 2026. Test from 7 December 2025 to 7 March 2026, across seven AI models. Otterly sells AI monitoring. https://otterly.ai/blog/schema-markup-real-impact-ai-search/ (accessed on 1 October 2026).
- OpenAI, Help Center, “Searching the web with ChatGPT”. https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt (read in the browser on 1 October 2026).
- Ahrefs, Louise Linehan, “Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)”, 12 December 2025. It is correlation. The sample is domains with Domain Rating above 40 and a keyword with at least 800 searches per month. https://ahrefs.com/blog/ai-brand-visibility-correlations/ (accessed on 1 October 2026).
- Otterly, Rick Tousseyn, “GEO Experiment: Your Image Alt Text is Built for SEO. AI Search Reads Right Past It”, updated on 21 April 2026. One week, on a test domain, across five AI models. https://otterly.ai/blog/geo-experiment-image-metadata/ (accessed on 1 October 2026).
- Ahrefs, Ryan Law, “New Study: AI Assistants Prefer to Cite “Fresher” Content (17 Million Citations Analyzed)”, 28 July 2025. https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content/ (accessed on 1 October 2026).
- Google Search Central, “Influence your byline dates in Google Search”. https://developers.google.com/search/docs/appearance/publication-dates (accessed on 1 October 2026).
- Mindo, guide written by murmur, “Vídeo para treinamento corporativo”, 2 June 2026. https://guia.mindo.com.br/video-para-treinamento-corporativo/ (accessed without running JavaScript on 1 October 2026).
- murmur, free diagnostic: https://murmur.marketing/en/diagnostic/; sector studies: https://murmur.marketing/estudos/ (accessed on 1 October 2026).