Updated on 19 October 2026, by João Leite, Founder of murmur.

AI citability is the chance of a page appearing in ChatGPT and the other AI models as a source, used to write the answer. The term circulates among GEO agencies, the work of making the company appear in the answers of ChatGPT and the other AI models. People looking for the service talk about appearing in ChatGPT or being cited by the AI. The published studies link citation to three things. The page covers the subject of the question, answers right at the beginning and brings proof that can be checked, such as a number, a definition and a comparison.

  • Citability belongs to the page, and presence belongs to the brand. Presence counts how many answers the brand appears in, with or without one of its pages cited.
  • The subject of the question comes before everything else. In a test with six AI models and 252 thousand rounds, the source outside the subject lost in all six.
  • Proof inside the text carries weight. In a study with 18,151 cited pages, those that carried definitions contributed, on average, 57.33% more to the answer.
  • Clear structure is part of the most supported recommendation. Martinez’s review of 45 studies asks for the page to be on the subject, complete, verifiable, with clear sections and findable by search.
  • The information gain comes from a Google patent, granted in 2022. The patent describes a method for choosing documents, and does not measure AI citation.

AI citability is the chance of the page appearing in ChatGPT as the source of the answer

AI citability is the chance of a page being chosen as a source and used by the AI to write the answer. That chance has two moments, which Zhang Kai and two colleagues separated in an April 2026 study. In the first, the AI searches and puts the page among the sources. In the second, the page delivers words, numbers or structure to the answer. The study looked at 18,151 pages cited by ChatGPT, by Perplexity and by Google’s AI Overview, or Google AIO, the AI-generated text that opens the results page in many Google searches.

Each AI splits the answer across a different number of sources. In the study by Zhang Kai and colleagues, ChatGPT cited an average of 6.88 sources per question, Google’s AI Overview 12.06 and Perplexity 16.35. ChatGPT cited fewer pages and made more use of each one of them in the answer.

The definitions of citability that circulate in the market use the term in different senses: the chance of a page becoming a source, the chance of a company being mentioned and even an editorial review criterion, without a number. The company’s sense already has another name, presence. Presence is how many answers the brand appears in, divided by the total number of answers. A citable page helps the brand appear, and the presence calculation shows whether it appeared.

murmur, a GEO agency in Brazil that places brands in the answers of ChatGPT, Claude, Google AIO and other AI agents, works with both calculations. It studies the client’s sector, writes the pages that answer the questions of that sector and tracks, week after week, in how many answers the brand appears. The page is what the AI can use as a source. Presence shows whether the brand entered the answers.

The factors measured to appear in ChatGPT as a source: subject, answer at the beginning and proof

The studies support three factors: the page covers the exact subject of the question, answers right at the beginning and brings proof in the text. The subject comes before everything else. Vishwakarma and colleagues, in a test presented at SIGIR 2026, an information retrieval conference, put six AI models to choose between two nearly identical sources, in 252 thousand rounds. The source outside the subject of the question lost in all six models, as did the source without a price, the one with an old date against the recent one and the one that came further down the list.

The sources in the SIGIR test were product reviews assembled for the experiment and delivered straight to the models, without ChatGPT’s search in between. The authors work at Sprinklr, a software company. Olivier Martinez’s review of 45 studies, from July 2026, gives strong support, in controlled tests, to the page that answers the question, alongside its position in the list of sources.

The answer at the beginning also shows up outside the laboratory. Surfer, which sells a text editor for SEO, scored the sources of more than 650 thousand real AI answers, in rounds from October 2025 and April 2026. In ChatGPT in 2026, the factors most linked to citation were in the first lines: the page that repeats the question and already delivers the short answer. Surfer writes that it is correlation, that is, it shows what the cited page looks like, and not why it was chosen.

The proof inside the text is the third factor. In the study by Zhang Kai and colleagues, pages with definitions contributed to the answer, on average, 57.33% more than pages without them, and those with numbers, 61.55% more. The authors treat these numbers as description, and not as cause. In the study by Aggarwal and others, which gave GEO its name and was presented at KDD 2024, putting sentences in quotes from sources on the page took the part of the answer attributed to it from 19.3% to 27.2%. There were a thousand questions, and in each one GPT-3.5, an OpenAI model, received the first five Google pages, one of them rewritten.

Martinez brings the three factors together in a recommendation that he himself calls conservative:

“The best-supported recommendation is conservative: produce a relevant, comprehensive, verifiable, clearly structured, and technically retrievable page; then measure retrieval, citation, and fidelity separately.”

In plain terms, Martinez’s recommendation asks for a page that addresses the question, covers the subject, makes each statement verifiable, has clear sections and can be found by search. Then each step is measured separately. The writing of that page, sentence by sentence, has its own guide: how to write a text to appear in ChatGPT.

Content format: the “best” list remains strong in “which is the best” questions, and clear structure serves both the reader and the AI

The “best” list remains strong in “which is the best” questions, although it has lost weight in ChatGPT in 2026. Ahrefs, which sells SEO and AI tracking tools, classified the types of page that ChatGPT cited in 750 early-purchase questions, partly by hand and partly with GPT-5. The questions were about software, products and agencies, in a study published in December 2025. “Best X” lists published on blogs, added to those the brands themselves publish, were 43.8% of the types of page cited. The number shows what ChatGPT considered and used in this type of question at the end of 2025.

In 2026, the list lost weight in ChatGPT. In July 2026, it was 9.7% of its citations, against 32.8% for the product page, according to Promptwatch, a company that sells AI measurement, in September 2026 and without the sample published. With the GPT-5.6 model, the list fell from 15.47% to 8.19% of the pages ChatGPT searches in a purchase question, according to Peec, which also sells AI measurement, in August 2026. Each calculation has its own sample, and none compares to Ahrefs’ 43.8%.

Clear structure is part of the page Martinez recommends, alongside the subject and the proof. The guide on how to write a text to appear in ChatGPT explains when to use a table, a list and a summary.

Chunking: cutting the page into passages, and the paragraph that stands on its own

Chunking is cutting a page into smaller passages, among which an AI search system chooses what to use in the answer. Anthropic, the company that makes Claude, describes the cut as the first step of a search base for AI. Documents become passages of, in general, at most a few hundred tokens, the pieces of words the model reads. In Anthropic’s own evaluation, the passage that loses the context of the document is found with more difficulty, and returning that context reduced failures in finding the right passage among the top 20 by 35%, from 5.7% to 3.7%.

Chunking asks one thing of the writer: the paragraph that makes sense on its own, with the subject named right at the start. A paragraph that opens with “this” loses its meaning when it becomes an excerpt. OpenAI does not say whether ChatGPT cuts up the page, and the written rule that exists is Bing’s, the Microsoft search engine that OpenAI lists among ChatGPT’s search partners. In its Webmaster Guidelines, Bing asks for the essential information at the start of the page and one subject per page. It also writes that the page has a better chance of being chosen as the basis for the answer and cited when the content stands on its own: “URLs are more likely to be selected for grounding queries and citations when content stands on its own”. The guide on how to write a text to appear in ChatGPT explains how to write that way, and the guide on how ChatGPT search works shows the path from the question to the sources.

Information gain: the text that adds to what has already been published

Information gain is how much new information a text adds to what has already been published on the subject. In Portuguese, ganho de informação. Some agencies that use the term point to a Google patent, “Contextual estimation of link information gain”, filed on 18 October 2018 and granted on 7 June 2022. The patent describes a score, calculated by a model, for how much a document brings beyond what the person has already seen, and the use of that score to choose what to show next.

A patent records a method, and does not show that Google uses it or that ChatGPT chooses sources that way. The “up to 40%” that sometimes appears next to the term comes from the Aggarwal study and measures something else: adding statistics, citations and sources increased the page’s space in the AI answer.

What helps with AI citation, and who measured each factor

These are the citability factors that the studies support, with who measured each one.

factorwhat was foundwho measured it, and how
subject of the questionthe source off the subject lost in all six modelsVishwakarma and colleagues, laboratory with six models
answer at the startthe factors in the first lines led in ChatGPTSurfer, correlation across more than 650 thousand answers
number, definition and comparisonthe pages with them contributed more to the answerZhang Kai and colleagues, 18,151 pages; Aggarwal, one thousand questions
recent datethe recent date beat the old one in all six modelsVishwakarma and colleagues, laboratory
clear structureappears in the most supported recommendation of the reviewMartinez, review of 45 studies

The page on the subject, with the answer at the start and the proof inside, is the one the studies support, and clear structure comes in as part of the content that the reader and the AI make use of.

What the page alone does not solve: the bot, the search and the other sites

Citability only counts for the page that the AI can find and read. Martinez writes that “a perfectly formulated page that is absent from the top-k cannot contribute”: the perfect page that stays outside the top search results does not reach the answer. AI crawlers, such as OpenAI’s GPTBot, are what each AI company sends to the site to download the text of the pages. They need to have their entry allowed and to receive that text ready in the HTML. Why text assembled by JavaScript may not reach the AI is in the guide on llms.txt, and the technical items, one by one, in the technical checklist to appear in ChatGPT.

What other sites say about the company affects the recommendation, and that is another job. Ahrefs cross-referenced 75 thousand brands in December 2025, and mentions of the brand on other sites accompanied its appearance in AI answers with a correlation of 0.66 to 0.71. Ahrefs warns that correlation does not prove cause. The reasons are in the guide on why AI recommends some companies and not others, and the weight of outside links is in the guide on backlinks to appear in AI. The citable page is the basis of the work, not the whole of it: what changes from an ordinary blog to GEO is in the guide on GEO and content marketing.

At murmur, the sector study chooses the pages, and presence shows whether they made it into the answers

At murmur, the client’s sector study decides which questions get a page, and ongoing monitoring of presence shows whether the brand made it into the answers. A specialist dedicated to the brand, the same person from the diagnostic onward, studies the company and the sector and surveys who appears in its place on each question. murmur writes the pages that fill the client’s gaps and publishes them on a subdomain of the client’s brand. If the contract ends, the pages stay with the client.

murmur writes these pages based on the academic studies of AI citation, among them Martinez’s review of 45 studies: the question answered at the start, with the proof inside the text. The result can be seen in the answers. murmur asks ChatGPT, Claude, Perplexity and Google AIO the same questions as the market, week after week, and the dashboard included in the service shows the brand’s presence alongside that of each competitor. Each murmur case study carries the client’s name, published with their authorization.

Frequently asked questions

Is citability the same as brand presence in AI models?

No: citability is the chance of a page becoming a source, and presence is the brand’s count in the answers. A cited page can make it into the answer without the company’s name, and the brand can be recommended because of what other sites say about it. The difference between being cited as a source and being recommended is in the guide on why AI recommends some companies and not others.

Can you measure the citability of a page?

You can measure the result, not a score: ask the AI models the market’s questions and count in how many answers the page becomes a source. The citability scores circulating in the market are methods of whoever created them, and no citation study that murmur has read links them to the choice of source. In Google, Search Console has a generative AI performance report, and what it shows and what it leaves out is in the guide on Search Console and Bing Webmaster Tools. For ChatGPT and the other AI models, the step-by-step is in the guide on how to measure brand presence in ChatGPT.

Is citability the same as SEO?

SEO and citability measure different things. SEO works on the position in Google search, and citability is the chance of the page becoming a source in an AI answer. murmur writes the pages that answer the market’s questions and tracks, in ChatGPT, Claude, Perplexity and Google AIO, in how many answers the company appears. Position in Google links is SEO work.

Does changing the page date make it get cited more?

The recent date beat the old one in a laboratory test, and it only counts when the text has actually changed. The test, and why the date has to be true, are in the guide on writing to appear in ChatGPT, in the question about changing the text date.

Where to start checking whether your pages become a source

The first step is to see what the AI models answer today to your market’s questions. murmur’s free diagnostic shows those answers for your company, with no sign-up and in less than two minutes. Each of murmur’s sector studies shows, in a niche sector, the brands that the AI models recommended in a round of 25 buying questions.

Read also

References

  • 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 and 18,151 pages downloaded, using the public data from geo-citation-lab; the figures for definition, comparison, step by step and question and answer are in the evidence genre table, and the word figures are in Figure 5. https://arxiv.org/abs/2604.25707 (accessed on 8 October 2026).
  • Rahul Vishwakarma, Shushant Kumar and Ratnesh Jamidar, Sprinklr, “What Gets Cited: Competitive GEO in AI Answer Engines”, SIGIR 2026, arXiv 2605.25517, 25 May 2026. Six models via API, 252,000 runs, 18 factors; the effects by model are in Table 2. https://arxiv.org/abs/2605.25517 (accessed on 8 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 factor is in Table 4. https://arxiv.org/abs/2607.14035 (accessed on 8 October 2026).
  • Surfer, Paulina Kaleta, “How to Rank in ChatGPT: What the Data Actually Says”, updated on 17 September 2026. More than 650 thousand AI answers collected from the tools’ screens, in the October 2025 and April 2026 rounds; 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 8 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 word count adjusted by position, in the test portion of one thousand questions from a bank of 10 thousand, with GPT-3.5 turbo. https://arxiv.org/abs/2311.09735 (accessed on 8 October 2026).
  • Ahrefs, Glen Allsopp, “Do Self-Promotional “Best” Lists Boost ChatGPT Visibility? Study of 26,283 Source URLs”, 4 December 2025. 750 early-stage purchase questions to ChatGPT, in software, products and agencies. Ahrefs sells SEO and AI monitoring tools. https://ahrefs.com/blog/best-lists-research/ (accessed on 8 October 2026).
  • Promptwatch, “Are Listicles Still Relevant in 2026? What AI Citation Data Actually Shows”, 9 September 2026. No sample published; Promptwatch sells AI measurement. https://promptwatch.com/blog/are-listicles-still-relevant (accessed on 15 October 2026).
  • Peec, David Konitzny, summary of an analysis on product pages and lists with GPT-5.6, August 2026. No sample published; the figure is of what ChatGPT searches for in a purchase question, not of what it cites. Peec sells AI measurement. https://docs.peec.ai/research/gpt-56-product-pages-vs-listicles (accessed on 13 October 2026).
  • Anthropic, “Introducing Contextual Retrieval”, 19 September 2024. Anthropic’s internal evaluation of its indexing method. https://www.anthropic.com/engineering/contextual-retrieval (accessed on 8 October 2026).
  • Microsoft, “Bing Webmaster Guidelines”, Bing Webmaster Tools, undated page. https://www.bing.com/webmasters/help/webmaster-guidelines-30fba23a (accessed on 1 October 2026).
  • OpenAI, “Searching the web with ChatGPT”, OpenAI Help Center, no exact date on the page. Lists Microsoft among ChatGPT’s search providers. https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt (accessed on 3 October 2026).
  • Google LLC, Victor Carbune and Pedro Gonnet Anders, patent US11354342B2, “Contextual estimation of link information gain”, filed on 18 October 2018 and granted on 7 June 2022. https://patents.google.com/patent/US11354342B2/en (accessed on 8 October 2026).
  • Ahrefs, Louise Linehan, “Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)”, 12 December 2025. It is a correlation. https://ahrefs.com/blog/ai-brand-visibility-correlations/ (accessed on 8 October 2026).
  • Google, “Generative AI performance report (Search)”, Search Console Help, with the note on the launch for all sites on 31 August 2026. https://support.google.com/webmasters/answer/16984139 (accessed on 8 October 2026).
  • murmur, free diagnostic: https://murmur.marketing/en/diagnostic/; studies by sector: https://murmur.marketing/estudos/; case studies published with client authorization: https://murmur.marketing/en/cases/ (accessed on 8 October 2026).