--- language: en title: "Where do sources cited by AI come from? How they differ by sector and engine" description: "An analysis of Profound’s study of 11.84 billion citations. See why company, editorial, and community sources need to be evaluated in context." canonical: "https://tropk.ai/en/resources/research/where-do-sources-cited-by-ai-come-from-how-they-differ-by-sector-and-engine" last_modified: "2026-10-12T16:14:21.281Z" --- # Where do sources cited by AI come from? How they differ by sector and engine Reviewed on 2026-10-06. ## The research and the decision it helps improve This editorial analysis was reviewed by Tropk on October 6, 2026, and refers to the Profound study published on July 30. The investigation is Profound’s, including its data collection, classification, and results. It shows that there is no single source ecosystem for all [AI searches](/en/resources/glossary/ai-search). The weight of company websites, publications, and communities varies by subject and engine. The practical marketing implication: deciding where to improve information requires looking at the sources that support audience questions. A strategy based only on the impression that AI prefers social media may overlook important documentation. The opposite idea—that publishing on your own domain is enough—may overlook the reviews and external references needed for comparison. ## Key points - Profound analyzed 11.84 billion citations across eight engines between April 16 and July 16, 2026. - The weight of company, editorial, and community sources varies by sector and engine. - The class of company-operated websites includes competitors and other companies, not just the monitored brand’s domain. - Global percentages and category medians use different units of analysis and are not targets for Brazil. - Classify the owner, page function, and supported facts before deciding where to improve information. ## What was counted and how the results were grouped The [original study](https://www.tryprofound.com/blog/where-do-ai-citations-come-from) analyzes 11.84 billion citations between April 16 and July 16, 2026, across eight engines. The dataset contains 8,061 active categories; 7,542 had at least one hundred citations. A map of 3.02 million domains covers 98.3% of the volume. Sources are classified as company-operated websites, editorial or institutional sources, and social or user-generated content platforms. There is also unclassified material. The industry breakdown uses category medians, subject to minimum volume criteria. This differs from adding up all citations and calculating a global proportion. A small category can carry the same weight as a much larger one in a median. Before using a percentage in your own report, preserve the unit of analysis. ## The findings do not mean your own domain dominates in every situation Profound reports that approximately 57% of the global volume comes from company-operated sources, and that this class predominates in the median for 24 of 29 sectors. The editorial or institutional share varies: the reported median is 11.4% for SaaS and 59% for pharmaceuticals. See the results in the [primary source](https://www.tryprofound.com/blog/where-do-ai-citations-come-from). The company class includes competitors and other companies, not just the monitored brand’s website. That last point changes the interpretation. If an AI cites ten company websites and none is yours, the result does not show that your own presence is well established. Nor does it show that all those references are commercial materials: a company may publish useful documentation, data, and explanations. Domain origin and page function are different dimensions. ## How to turn the classification into a useful diagnosis To analyze your own market, first record the URL and the fact it supports. Then classify the domain owner and the content’s role separately. An independent review, a specifications page, and a user discussion may cover the same product while serving different purposes. This approach avoids treating every external reference as a PR opportunity. An operational classification can distinguish your company, other companies, publications, institutions, and communities. This would be a team’s editorial adaptation, not a direct reproduction of the study’s three classes. Document that difference to avoid comparing incompatible percentages. In ambiguous cases, keep the URL and a review note rather than forcing it into a category that only looks precise in a chart. Volume also needs context. Several references to the same domain in one answer may support different facts; counting unique domains alone answers a different question. Choose the rule before collecting data. If the goal is source presence in answers, count the answers that include that source. If the goal is reference diversity, keep URLs and explain the normalization. ## What this means for a team in Brazil A company selling in Brazil should assess whether sources cover availability, terms, and criteria relevant to Brazilian buyers. A global comparison may be informative yet omit Portuguese-language support or a local integration. The action may involve your own website, a partner’s documentation, or an incorrect external listing; it does not have to start with a new generic article. The study does not provide a representative breakdown for Brazil or Portuguese. Therefore, the 57% global share is neither a local target nor a budget recommendation. To choose between documentation, media relations, and community participation, compare actual gaps, audience reach, and your ability to keep information accurate. Citation frequency is a signal, not the only investment criterion. Also consider whether the source suits the type of claim. For product specifications, official documentation may be necessary. For user experience, accounts can add context. For a comparison, clear criteria and verifiable data matter more than simply appearing on a frequently [cited domain](/en/resources/glossary/cited-domain). Copying another sector’s source distribution does not ensure this fit. ## A local survey before choosing channels As a proposed exercise, gather twenty answers to important questions across discovery, comparison, and validation. This sample size is a starting point for manual inspection; it does not come from the study and does not represent the market. Record the engine, location, question, brands, URLs, and facts. Repeat some questions to identify variation and avoid mistaking a single answer for a stable pattern. Look for actionable problems: a nonexistent feature attributed to the brand, a competitor cited as a category reference, a question no page answers, or an outdated review that is frequently used. For each problem, identify who can correct the information and where. Plan the action around an observable outcome, such as providing an accurate specification, rather than promising to win a share of citations. ## Limitations to include in any presentation of the data - The dataset reflects categories monitored on Profound’s platform, not a census of all consumer conversations. - Classification depends on a domain map and taxonomy decisions. Websites may have mixed functions, change ownership, or be misclassified. - The observed period is specific. The distribution may change with models, interfaces, categories, and available sources; it does not demonstrate the causal effect of a marketing action. ## Reading and next steps Use [primary source](/en/resources/glossary/primary-source) and [AI citation](/en/resources/glossary/ai-citation) to check what each reference actually substantiates. The [research and benchmarking prompt](/en/resources/prompts/prompt-for-writing-a-research-study-or-benchmark) helps make the sample, criteria, and limitations explicit. [Market intelligence](/en/product/market-intelligence) and [Strategy](/en/product/strategy) connect this survey to content and distribution choices. Priorities should follow from the observed problem and the buyer, with the original source preserved. ## Sources and references - [Where do AI citations come from?]() — Profound — Jasman Singh ## Frequently asked questions ### What did Profound’s study analyze about the origins of sources cited by AI? Profound analyzed 11.84 billion citations across eight engines between April 16 and July 16, 2026, classifying sources as company, editorial or institutional, and social or user-generated. The distribution varies by subject and engine. The dataset reflects categories monitored by the platform, not a census of all consumer conversations. ### Does the approximately 57% share of company sources mean that publishing on your own website is enough to be cited by AI? No. In Profound’s study, the company source class includes competitors’ and other companies’ websites, not only the analyzed brand’s domain. Domain origin alone also does not tell you a page’s function. Check which URLs support the answers and what facts they provide for the buyer’s needs. ### Why is a category median not the same as the global share of AI citations? A global share adds up citations and calculates each class’s proportion of the total volume. A median summarizes the percentages across categories and can give similar weight to categories with different volumes. Profound’s study uses these units in different presentations; comparing them as if they were the same measure leads to incorrect conclusions. ### How can research on sources cited by AI inform marketing decisions in Brazil? Inspect questions relevant to Brazilian buyers, record URLs and supported facts, and classify the domain owner and content function separately. Check availability, local terms, and information gaps before choosing an intervention channel. Profound’s global percentages are not Brazilian benchmarks and do not demonstrate the causal effect of publishing in a particular type of source. ## Continue a leitura - [Topical authority]() - [AI citation]() - [Primary source]() - [Search intent]() - [Prompt for writing a research study or benchmark]() - [Market intelligence]() - [AI discovery strategy]() ## About the author [Luiz Gardelin]() Luiz Gardelin is the founder of Tropk.ai and works in GEO/AEO, product vision, strategy, and growth, helping brands understand and improve their presence in AI responses. --- [All pages in Markdown]()