Methodology — version 1.0

We measure presence: whether, how often and in what way an entity (a business, a practice, a professional, a news outlet, a journalist, a creator) is named in the answers that AI engines give to questions asked by a generic user in the market a report covers. A "mention" is a recognizable reference to the entity in the text of an answer; sources (the links the engine attaches) are recorded separately.

Rankings measure presence in the answers of AI engines. They do not confer titles, credentials or any judgment of professional quality.

The distinction between ChatGPT and ChatGPT Search matters. The first shows what the model learned during training; the second shows what it finds on the web at the moment of the question. An entity that appears only in the search-enabled modes is visible but not yet learned by the model; an entity named by the model alone has entered its knowledge. This is why the two modes are counted separately.

If an engine cannot be reached, or delivers no answers during the collection period, the report says so, with the number of missing answers. Missing data is never estimated or interpolated.

Each report uses a fixed set of prompts, designed to mirror the way people actually ask an AI assistant for a recommendation: by category, by city, as a direct request and as a concrete problem. The number of prompts and their complete list are part of the report's frozen protocol, and are published in the dataset before collection begins.

The prompts:

• are written in the language of the market the report covers, from the point of view of a generic user (no personal account, no history);
• never name a specific entity (no "tell me about Business X" in a ranking collection);
• stay identical from one report to the next within the same series; any change produces a new version of the methodology, published before the report that uses it.

Rankings of news outlets, journalists and creators

Some reports rank the outlets, journalists and creators of a field (for example artificial intelligence) instead of businesses. For these reports:

• each category uses one plural question, for example "Who are the most authoritative journalists covering artificial intelligence in the United States?", repeated many times rather than reworded many ways;
• the question never asks for a number of names: the engine answers with the list it chooses, because asking for a fixed number pushes engines to fill the list with invented names;
• when a question uses a word such as "authoritative", the result is always attributed to the engines: a ranking shows whom AI engines name when they are asked that question. It is not a judgment by Visibilitas, and we do not measure authority itself.

• The professional quality, competence or results of the entities named.
• Titles, credentials or specializations: presence in AI answers is not a title.
• Whether what the engines say is correct: we record what they answer, we do not certify it. Errors found in the answers are noted in the dataset.
• The traffic, inquiries or customers that mentions generate.

1. Engine volatility. AI systems change models and search mechanisms without notice: two reports far apart in time are not fully comparable, and comparisons over time should be read as trends, not as absolute measures.
2. Sampling. A fixed set of prompts does not exhaust the ways people ask; it measures a fixed, declared sample of representative wordings.
3. Collection infrastructure. Part of the collection runs through specialized third-party platforms. Malfunctions, delays or missing engines are stated in the report; the raw data stays in our hands and is published in full, and the collected data is checked continuously.
4. Non-determinism. The same prompt can produce different answers at different times: repetition across days, prompts and engines exists precisely to reduce the weight of any single answer.