No ranking without data
If the sample is not enough, the edition is not published. There are no estimated, filled-in or reconstructed values.
Observatory method
Every ranking, step by step.
Every ranking starts from published questions, repeated on several AI assistants, read with fixed rules and counted with their margin of uncertainty. Here you will find the steps, the thresholds, the sources and the limits.
The cycle
Eight steps, always the same.
The same cycle applies to every series and every country. Only the source of the topics changes. Weekly rankings are published every Monday, on the week just closed.
The signal
An external source tells what matters at that moment. It is always an official or licensed API or dataset, never a copied page.
Choosing the topics
The editorial team keeps the topics where a recommendation makes sense and discards breaking news, private individuals and sensitive topics. Choices and discards stay on record.
The questions
For each topic, three fixed phrasings (choice, comparison, budget), written in the country’s language and published word for word.
Collection
Each question goes to each assistant, from an access point in the country, in fresh sessions. From three to ten repetitions, until the interval is narrow enough.
Reading
A model suggests the names present in the answer; the code accepts them only if they really appear in the text. New or ambiguous cases go to a person.
The count
For each brand, the share of answers in which it is named first and the share in which it is named, with a 95% interval.
The thresholds
At least three valid answers per cell, at least 70% of cells measured, at least eight topics. Below the threshold the edition is not published, and we say so.
Closing
The edition is frozen with date, sample, assistants and files. It is never recalculated: an error leads to a public withdrawal and a correction next to it.
Assistants and channels
Each assistant, its own channel.
An API and an interface can answer the same question differently. That is why each assistant’s channel is declared, and data from different channels is never mixed. Each edition lists the assistants measured.
| Assistant | Channel | Web search |
|---|---|---|
| ChatGPT | Web interface and OpenAI API, in separate columns | Yes |
| Gemini | Google Gemini API | Yes, with Google Search |
| Perplexity | Perplexity API and web interface, in separate columns | Yes |
| Claude | Anthropic API | Yes |
| Copilot | Web interface | Yes |
| AI Overviews | Google results from a licensed provider | Part of the search |
| AI Mode | Web interface | Yes |
The definitions
Every word, one measure.
The words of the rankings have a precise meaning, the same in every edition.
- Named first
- The first brand in the registry that appears in the answer text. It is the measure that orders the rankings.
- Named
- The brand appears at least once in the answer, in any position.
- Recommended
- The answer points to it as a recommended choice, not just as an example. This is decided by rules and review, never by the model alone.
- Cited
- The answer links to a page on the brand’s domain. It counts only where the assistant shows its sources.
- Share of voice
- The brand’s mentions out of all brand mentions in the series’ answers.
- Tie
- Two brands whose intervals overlap. The table says so instead of inventing an order.
Topic sources
Topics only from official channels.
No page is copied automatically. Every reading of the signal carries its source, time and fingerprint, so a topic can always be traced back to its origin.
| Series | Signal source | Type |
|---|---|---|
| Trending topics | Google Trends | Official API |
| The most searched questions | Editorial questions by sector, ordered with Google Trends and Wikimedia Pageviews | Editorial list and official APIs |
| Before buying | Editorial purchase categories, ordered with Google Trends | Editorial list and official API |
| Leading brands | Wikidata | CC0 open dataset |
| Who makes the press | Webz.io News Search and GDELT | Licensed API and open dataset |
What we don’t do
The rules that do not bend.
Four commitments apply to every series, every country and every edition.
No number from the model
The model reads the texts and suggests names. Counting, weighting and ordering are done by code, with published rules.
No judgement on products
We report what AI assistants answered. We do not say which brand is better.
No positions for sale
The order does not depend on subscriptions, sponsorships or requests from brands.
Reusing the data
Open data, under a CC BY 4.0 licence.
Rankings and counts for each edition can be downloaded as CSV and JSON with the field dictionary and the method version. You can reuse them, including for commercial purposes, by citing “TrustInsider, Observatory”, the series, the country and the edition.
A sample sentence written by someone reusing an Observatory data point: the notes show the elements the licence requires you to cite.
QuestionHow do you cite an Observatory data point in an article?
Sample article reusing the data
According to “TrustInsider, Observatory”, Trending topics series, Italy, edition for the stated week, the brand cited was named first more often than the others.
Source
The attribution required by the CC BY 4.0 licence, word for word.
Citation
The series and the country, so that the data point can be found in the right file.
Citation
The edition, because closed editions are not recalculated.
Errors and corrections
An error is disclosed.
Anyone who believes a figure is wrong writes to the editorial team stating the edition and the row. The check starts again from the recorded answers; the outcome and any correction stay public next to the edition.
The path of an error report on an edition: from the request to the editorial team to the public note next to the edition.
Write to the editorial team
You state the edition and the row you believe is wrong.
We verify
The check starts again from the recorded answers, not from an estimate.
We publish the outcome
The outcome stays public next to the edition, even if the data point holds.
We correct if needed
The incorrect edition is withdrawn with a note, and the correction is published alongside.
Frequently asked questions
On the method, in brief.
How many times we repeat each question, which confidence interval we use, how we weight the assistants and why questions start from accounts with no history.
How many times do you repeat each question?
Three to ten times per question and assistant, until the interval is narrow enough: a single answer is an anecdote, several answers give a frequency.
Which interval do you use?
The 95% Jeffreys interval. With few samples per cell it is more reliable than the normal approximation and always stays between 0 and 100%. It is the same method used by the TrustInsider platform.
How do you weight the assistants?
By each assistant’s usage share in the country, from public sources that are declared and dated in the edition note. The ranking for each single assistant always stays visible alongside.
Are the questions sent from an account with a history?
No. We use fresh sessions, with no history or preferences. It is a laboratory condition: a real person may see different answers.
Related paths
After “Observatory method”, the next step.
The Observatory publishes rankings with their methodology. From here you move on to the other series, to the edition methodology and to the page that brings them all together.
Who comes first in AI answers.
Regular rankings of which brands AI assistants name first, in six countries, with a public method and open data under a CC BY 4.0 licence.
ReadA topic is rising. Who does AI name first?
Every week, the topics rising in searches and the brands AI assistants name first, country by country.
ReadThe usual questions. Today’s recommendations.
Every month, the questions customers ask before choosing and the brands AI assistants recommend, sector by sector.
ReadBefore buying. Who the assistant recommends.
Every month, the brands and stores AI assistants suggest for the purchase categories with the most interest.
ReadLeading brands. How much room in the answers.
Every month, how much room AI assistants give the leading brands of each sector and country, and with which sources.
ReadIn the press and in the answers. Do they match?
Every week and every month, the brands most present in articles, compared with what AI assistants answer.
Read