A mention occurs when a text names a brand without giving a source for it. For example, asked about time-tracking software for small businesses, an AI system names three providers, describes each in one sentence, and includes no links. Each of these three names is a mention. If an answer instead gives a page as a source—with a link, a footnote, or a source card—that is an AI citation. Both are part of a brand’s AI visibility, and they occur independently. If an answer names the brand and also cites its website, it contains both: a mention and an AI citation.
Mentions in AI answers
In an AI answer, a brand can come up as a recommendation, in a list, in a comparison, or as an example of a product category. For a mention, all that counts is that the brand is named. Whether the answer portrays it correctly is a separate question (brand accuracy).
Mentions are captured with a fixed prompt set, a collection of typical questions that prompt tracking puts to AI systems such as ChatGPT, Gemini, Claude, and Perplexity on a regular basis. A simple metric for this is mention rate, the share of answers in which a brand appears. Citations are counted separately, for example as citation rate, because a brand can be mentioned often while its website is rarely cited, or the other way around. Reporting both figures side by side shows whether AI systems bring up a brand on their own and whether they give its website as a source.
Mentions on other websites
Mentions also exist outside AI answers: in trade articles, product tests, comparisons, forums, directories, and reviews that name a brand, often without linking to its website. Such mentions can reach AI answers through a language model’s training data or through pages an AI system retrieves for an answer (grounding). What matters for mentions is that models learn from the text itself: a mention can enter their knowledge even without a link, and how well a model knows a brand likely also depends on how often the brand appears in its training data.
Coverage that others publish about a brand on their own initiative is called earned media, and it is an important source of such mentions. Seeking out inauthentic mentions, on the other hand, is not as helpful for the AI features in Google Search as it might seem, according to Google. How a brand builds genuine presence on other websites is the subject of Off-Page GEO. An overview of which other websites name a brand, and in what context, shows what material AI systems can draw on; compared with pages that name competitors, it also shows where the brand is missing.
Where mentions in AI answers come from
A brand appearing on a page that an answer cites does not mean the answer will name it. Nor can every brand an answer names be found in the sources it gives. A preprint from July 2026, a study released ahead of peer review, illustrates this. The researchers belong to the research institute of a Beijing technology company and collected their data with a tool for monitoring AI answers whose name the institute shares. Mostly in June and July 2026, they put 614 questions in Chinese to four Chinese AI systems—Doubao, DeepSeek, Tencent Yuanbao, and Qwen—via both web and app, three times each. They then compared the brands in the answers with the brands in the text of the cited sources. Of the brands that appeared in the sources, the answers named only about 8%. Conversely, about 13% of the brand mentions in the answers could not be matched to any brand in the sources cited at the same time.
The study left open where these mentions came from. Possible origins include the language model’s parametric knowledge, meaning what it learned in training, or sources the system does not display. The measurement itself may also have gaps: from the cited pages, the analysis captured only the main text—without buttons, contact details in footers, or structured data—and pages may have changed by the time they were analyzed. Whether the figures carry over to ChatGPT, Gemini, Claude, and Perplexity was not studied. The direction is still telling: mentions do not simply follow the sources an answer gives. They can therefore only be captured by analyzing the answer text itself.
Recognizing name variants
To count mentions, an analysis first has to find them in the text. That is harder than it sounds, because a brand appears in many forms: with and without its legal form (“Miller Plumbing LLC,” “Miller Plumbing”), with or without an apostrophe (“Miller’s Plumbing”), abbreviated (“MP”), under a former name, or simply misspelled. In language processing, the task of finding names in a text and classifying them by type, such as person, place, organization, or product, is called named entity recognition.
Analyses of AI answers take different approaches. Some search the text for a fixed list of brand names; researchers at the University of St. Gallen point out in an April 2026 preprint that such matching misses brands named through synonyms, abbreviations, or paraphrases. Others have a language model extract the brands named in the answer text, as researchers at the University of Toronto did in a September 2025 preprint. Even then, the different spellings found have to be assigned to the same brand afterward. Every reliable count therefore rests on as complete a list as possible of a brand’s names—with legal form, short forms, former names, and common misspellings—used the same way in every measurement.