AI Brand CheckWe monitor how frequently your company appears in responses to selected prompts. Each key figure is based on saved answers and a defined counting rule.
The current audit uses OpenAI with web search via the API. It is a snapshot of selected questions, not a direct measurement of all ChatGPT users. Answers in the ChatGPT app may vary.
An observation is a response to a specific prompt in a specified language, with a specified country specification and a repetition. Every request starts without a previous conversation. We store original answer, citations, other sources accessed, query parameters, model and time.
New suggestions explicitly ask for suitable providers. We use this to measure visibility when searching for a provider, not for all types of information questions. Your own changes can change this comparison basis. Previous reports and manually edited prompts will not be rewritten retroactively.
The standard package contains eight prompts and three repetitions in one country and one language: 24 answers. Questions with brand names are evaluated separately; The number of answers for the open provider search therefore depends on the prompts you choose. More countries, languages or repetitions increase the measurement plan.
Q: Prompts · R: Repetitions · C: Countries · Lc: different selected languages in the respective country.
New audits cover four perspectives: location, service/product, purchase decision and brand mention. Suggestions reflect the documented offering; brand prompts include the business name. Categories are fixed groups, while their texts remain editable. You can add prompts to each group. Previous audits retain their original configuration.
Prompts with a company name, stored alias or domain are recognized as brand prompts regardless of the category. Your answers will appear in a separate report and CSV, but not in general visibility, competitive rankings, source analysis, or bootstrap intervals.
Live circles: Green means mentioned, red not mentioned, gray not clearly evaluable. Empty circles are pending runs. A prompt line is only displayed after the evaluation has been completed. The detailed view still distinguishes between recommendations and pure mentions. Red is not a quality rating for the company. A negative statement can occur at the same time as a recommendation and is reported separately.
The most common competitor per prompt is determined based on the number of documented recommendations on a common, evaluable response basis. Each provider counts once per answer; Ties are shown together. CSV export contains one line per scheduled run. Missing values remain unclear. To protect against table formulas, text fields that begin accordingly receive a leading apostrophe; the saved original answers remain unchanged.
Mentions: The company name or a used alias appears as a full name phrase in the response text. Link texts count; a URL alone doesn't count. A mention can be positive, neutral or critical.
Recommended: The company is suggested as a suitable option for the matter. The automatic classification requires an exact passage of text as evidence. A casual mention or mere citing of the source is not enough.
Website cited: At least one URL citation belongs to your own domain or a subdomain. Pages accessed in the background but not cited do not count.
Negatively mentioned: An explicitly critical statement with exact textual evidence. Not receiving a recommendation does not mean being mentioned negatively. Older runs without this evaluation remain marked as “not collected”.
Each brand and domain counts at most once per response. Ten mentions in one response do not count ten times. The three rates are independent and must not be added together.
A separate denominator is used for each key figure. An unclear recommendation can still contain a clearly recognizable mention and citation. Missing answers are not interpreted as non-mentions. The report states absolute frequencies and open ratings.
The semantic evaluation is automated. An exact quotation proves the basis of the text, but does not guarantee the correct interpretation. Same-named companies, ambiguous statements, and inconsistent aliases can cause errors. The original answer remains available for checking.
Specifically named providers classified as relevant are recorded. The order in the ranking depends on the selected rate. It is not a quality judgment about companies. For the ranking, only answers are used for each key figure in which both the provider evaluation and the observation of the company's own brand are known. All providers are counted on exactly this common response basis. The overview, on the other hand, uses all known observations per key figure; its denominator can therefore be larger. Excluded answers are reported. Ties receive the same rank: Rank = 1 + number of providers with a strictly higher number of hits. A rank is not a statement of statistical significance.
This share uses only the same set of responses with an available provider evaluation. Each brand counts once per response. It is an observed share of mentions, not a market share or a share of total AI search volume.
A competitive gap exists when an evaluable answer mentions other relevant providers but not your brand. Prioritization is based on the number of such responses; If there is a tie, the order of the prompts remains the same. Sales potential and search volume are not estimated.
Generative answers may vary. Repetitions examine this fluctuation while maintaining the same prompt. They do not replace a broader questionnaire or measurement at a later date.
Only repetitions of the same prompt, the same country and the same language are compared. With R valid answers, R(R−1)/2 pairs arise. Without an evaluable pair, stability remains open.
Under the simplifying assumption of independent answers with a constant recommendation probability p, the probability of never observing a possible recommendation in R trials is:
Assuming p = 20%, this is 51.2% for three and 32.8% for five repetitions. This illustrates the limitations of small samples; p is not a measured industry value. Common search results and model states can violate the independence assumption.
For fully evaluable repetitions, we draw as many recommendation decisions within each prompt with replacement as were made. From this we calculate the recommendation rate again. This is repeated 10,000 times, separated by country and language and with a fixed random starting value of 4127.
r* are the recalculated rates; Q denotes their empirical quantile. The code uses the sorted values at positions 250 and 9750. The interval describes the observed repeat variation under this measurement plan. It is not a guarantee of 95% coverage for small or dependent samples.
For three identical results, a bootstrap interval can be, for example, 0-0%. This does not prove the impossibility of a future recommendation. If the evaluation is incomplete, no interval is specified. This does not include selection bias, translation errors and systematic classification errors.
An initial audit should make differences between specific markets visible without querying each country individually. We use one for that targeted, non-random market selection. It is based on the concept of “judgment sampling”: suitable for exploration but susceptible to selection bias. This does not result in any statistical representativeness for Europe or the world. This limit also describes Statistics Canada.
| Package | Specific countries |
|---|---|
| Switzerland | Switzerland |
| DACH | Switzerland, Germany, Austria |
| Europe · Selection | Switzerland, Germany, France, Italy, United Kingdom, Poland |
| EU · Selection | Germany, France, Italy, Poland, Sweden |
| USA | United States |
| North America · Selection | United States, Canada, Mexico |
| South America · Selection | Brazil, Argentina, Chile |
| Asia · Selection | Japan, India, Singapore, United Arab Emirates |
| Africa · Selection | South Africa, Nigeria, Egypt |
| Oceania · Selection | Australia, New Zealand |
| Whole world · Selection | Switzerland, Germany, France, United States, Brazil, South Africa, Egypt, India, Japan, Singapore, Australia |
EU: Germany and France (west), Italy (south), Poland (east) and Sweden (north). North America: USA, Canada and Mexico, with English, French and Spanish language contexts. South America: Brazil for Portuguese and Argentina and Chile as two Spanish-speaking comparison markets. Asia: Japan, India, Singapore and the United Arab Emirates for East, South, Southeast and West Asian contexts. Africa: South Africa, Nigeria and Egypt for southern, western and northern sub-regions. Oceania: Australia and New Zealand; the Pacific island states are not taken into account.
Worldwide, the selection combines the Swiss starting point with Germany and France as well as the USA, Brazil, South Africa, Egypt, India, Japan, Singapore and Australia. This includes the five major geographical regions of Europe, America, Africa, Asia and Oceania. Many countries and sub-regions remain expressly untested.
The specific countries are chosen editorially, not through a validated optimization model. We do not use any claimed AI usage numbers, market shares or economic weights. Other countries within the same subregion may produce different results. For example, China, Central Asia and numerous African markets are not included in the global selection. The results may therefore only be based on the countries actually tested.
Your actual target markets are what matter: remove countries where you do not operate and add any missing markets. You can select countries individually; your final selection is saved with the audit. A location named in a prompt is retained in translation. A prompt about Aarau still tests a need in Aarau, even when Germany is set as the country.
Selection version: purposeful-markets-v2 · As of September 18, 2026. Changes to packages do not change previously saved audits. Repetitions do not increase geographical representativeness.
The overall score corresponds to the recommendation rate on a scale from 0 to 100: 44 points mean that the company was recommended in around 44% of the evaluated responses. It is not an additional weighted quality index. Mentions and website citations remain independent information. The preview for the example company AG illustrates solar installations and heat pumps for private households.
Country and language are separate settings. The country parameter describes an approximate user location; it does not guarantee a browser physically running there. For language comparisons, the country, prompts and number of repetitions should remain the same.
The current overall display adds up the evaluable answers of the selected filters. Combinations with more evaluable answers are given more weight. It is not an equally weighted country index. If there are different coverages, we recommend comparing individual countries and languages.
Domains and pages are ordered according to the number of answers with a citation. Page URLs are cleaned for grouping by fragments and the utm_source parameter; the original link remains.
A source in a response with competitors is an audit note. This does not mean that your brand is missing from this page, nor does it mean that the source caused the recommendation. Without additional page checking, we therefore speak of “sources for checking”. Competitor websites will not be classified as cited or uncited without confirmed domain attribution.
Each fully observed response belongs to exactly one of four groups: brand mentioned and own website cited; brand mentioned only; own website cited only; neither brand mentioned nor own website cited. Unknown values are kept separate. These groups do not overlap, and their sum equals the number of responses for which both observations are available.
Topic coverage counts prompts with at least one brand mention. It is not market coverage. The short conclusion and the diagnoses are rule-based and derived from the stored observations; no additional AI score is generated. There is no universal limit for “good” or “bad”. The competitive comparison exclusively describes the selected audit.
Prompts, individual repetitions, translations, complete answers and sources can be opened in the report. Original texts remain unchanged. The evaluation is versioned; New classifications are not implicitly incorporated into older measurements.
Reports are stored securely online. Once payment is confirmed, your analysis continues on the server even if you close your browser. A private access link lets you return without an account, including from another device. You can create and save it on the analysis page or in your report; it is valid for 90 days and can be revoked. No automatic email is sent. Save your access link or download the access file. If you lose it, our support team can help after verifying your order.
For quality assurance and further development of AI Brand Check, our team can view your submitted entries and results. Unsubmitted form entries will not be recorded.
OpenAI: Web Search, Sources and Location Preferences ↗
Switzerland, Europe and worldwide: three perspectives on AI in everyday life. We separate measured clicks from AI responses from usage surveys. This way you can see what the numbers say for your company.
Switzerland · Europe · Worldwide →Mobile lab measurement of a single page, not a full website scan. Technical quality is not a measurement of AI recommendations. Automatic tests do not prove complete accessibility.
Checks the technical operability by AI agents. No ranking and no guarantee that ChatGPT will recommend you. Inapplicable and untestable tests do not count as errors.
Measurement data: Google PageSpeed Insights / Lighthouse. Lighthouse: Agentic Browsing ↗