Method
The whole point of the index is that someone else can recompute it. Here is exactly how.
The questions
We use 38 open questions. Open means we never name a brand in the question, because we are measuring which brands the model puts forward on its own. Comparative questions of the kind "is X or Y better" measure something else and do not belong in the same index.
We never measure a sector in which the publisher has interests of its own. That rule holds even when the category would be the most relevant one for our readers.
The runs
Every question is put to every model 3 times. Models answer differently from one run to the next, so a single answer tells you nothing. We report frequency, never one answer. In total 456 answers were analysed in this measurement.
Web search
Every model runs with web search enabled where the provider supports it. That is closer to what a consumer actually meets than a model answering purely from training data.
A limitation we state openly: we measure through the providers' APIs, not through their chat interfaces. The chat products carry their own system prompts and post-processing, so the figures here are a close approximation of the consumer experience rather than an exact copy of it. Now that the chat products have begun showing ads, that matters in a new way: the APIs carry no paid placement, so what we measure is the unpaid layer. That is deliberate. It is the part a brand can influence without buying its way in. Scraping the chat interfaces would breach the providers' terms.
How brands are found in the text
By deterministic name matching against a register of brands and their aliases, not by a model. Matching happens on word boundaries, and ambiguous names require context. The operator "Tre" only counts when the word sits near terms like mobile, plan or coverage, since otherwise the Swedish numeral in "the three best" would register as a hit.
The Swedish genitive counts as a mention. "Telias bredband" names Telia just as much as "Telia" does.
Every month we review a list of capitalised names that no alias covered. That is how new brands enter the register.
When a brand gets its own page
Only once it is named across at least 3 different questions. Below that threshold there is too little to say, and the page would be thin auto-generated content.
Can a brand pay for placement?
No, and you do not have to take our word for it. The raw data from every measurement is downloadable under Open data together with the code that counts. An index whose inputs are public cannot be rigged.
The site carries advertising. It is always labelled, always sits outside the data tables, and never affects the figures, because the figures are computed mechanically from the raw data.