How the Index is measured
Which questions are asked, how often, how a name becomes a share, and when a page is ranked. Everything here applies to every page of the November 2026 edition.
The Index asks ChatGPT, Claude and Gemini four buyer questions per kind of business and place, three times each, without web search. A business is listed once two answers name it, ranked by its share of answers with a 95% interval, and a page is ranked only when enough answers came back.
What is asked
- Assistants
- ChatGPT, Claude and Gemini, asked in November 2026: ChatGPT and Gemini through a gateway to the maker’s API; Claude directly through the maker’s API. The free scan asks them too.
- Questions
- Four per kind of business, written out in the language of the place rather than translated word for word, and asked in the same order everywhere. Each ranking page prints them, with an English gloss on the English page.
- Runs
- Every question is asked three times to every assistant. A single answer is a sample; three show which names come back.
- Prompt
- The system prompt and recommendation format the free scan uses. The prompt is in English for every place, and the questions are in the place’s language. No web search, no location and no chat history, so the answer comes from what the model already holds. In a November 2026 test on 80 pages in five languages, asking in English instead of the local language barely changed how often Claude named no one, while a translated prompt changed it a lot, up in some languages and down in others.
- Places
- Local trades are asked about a city, software, online services and consumer brands such as banks, mobile carriers and supermarkets about a country.
What is counted
- AI Recommendation Rate
- The fraction of answers that named the business. A business named twice in one answer counts once.
- No name given
- An answer that names no business still counts as an answer, so it lowers every share on the page. Each page says how many answers from each assistant named no one.
- 95% interval
- A Wilson interval around the share. With a few dozen answers per page it is wide, and two businesses whose intervals overlap should be read as level.
- Listed
- A business appears in a ranking once two or more answers name it. Names given once are counted in the page totals but not listed.
- Order
- By share, then by average position in the answers that named it, then alphabetically.
- Named again on a rerun
- For each assistant and question, the names that came back in two or more runs, as a share of every name the runs gave.
- Overlap between assistants
- The average, over each pair of assistants that both named someone on the page, of the names both gave as a share of the names either gave. An assistant that named no one is left out of the pairs rather than counted as disagreeing.
When a page is ranked
4,939 of 4,948 pages met all three conditions this edition.
- Answers
- At least 24 answers across the assistants.
- Failures
- Fewer than 20% of the calls failed or came back unreadable.
- Names
- At least five businesses named in two or more answers.
- Below the line
- A page that misses any of the three is still published for anyone who has its address, marked as not ranked, left out of search engines and not linked from the Index.
How names are merged
- Spelling
- Names are compared with accents, capitals, legal suffixes, a leading "the" and notes in brackets removed. Nothing fuzzier: two real practices folded into one would be a worse error than one practice listed twice.
- Branches and chains
- A language model then groups the spellings on each page that are one business, such as a chain's branches, and is told to keep names apart when unsure. The groups are stored with the raw answers, and the other spellings are printed under the name.
- By hand
- A kept list of overrides can merge two names or keep them apart, and wins over the model.
- Not verified
- Names are printed as the assistants wrote them. We do not check that each business exists, is open or is any good. In a hand check of 180 names drawn at random from the November 2026 answers, 49% matched a business in that city and trade, 16% a real business in another city or trade and 4% one that has closed. For 9% we found nothing, and the rest could not be settled. The match rate was 67% on English pages, 40% on Italian ones and 42% on German ones.
What the Index is not
It is not a rating of the businesses, a local search result or a forecast. It records what these models said, in one month, to questions asked without a location or a chat history. A real conversation can go differently, and the next model version will.
No business can pay to appear in the Index, rise in it or be removed from it. How the Index stays neutral
Corrections and reuse
Archived for citation: DOI 10.5281/zenodo.23934866, Hugging Face, Kaggle
To cite this edition: AIGROW (2026). AI Visibility Index, November 2026 edition [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23934866
See what they say about you
The free scan puts your own customers' questions to the same assistants, with the same prompt.
Edition November 2026 · published by Guglielmo Paesano, founder of AIGROW · data under CC BY 4.0