How NZBrands.ai studies work

More people now ask an AI assistant before they choose a product, a service or a vote. A study asks the assistants the questions people actually ask, saves what they say, and reports it: who gets mentioned, what is said about them, and where the information comes from.

We report what the assistants said. We take no view on any business, party or policy, and no one pays to be included or left out.

We start with people, not topics. For each study we describe the situations of the people it is about, then write the questions someone in each situation would plausibly ask, including their circumstances when they would mention them. The wording is plain and does not lead the assistant towards any answer.

The people are designed situations. Their priorities are our assumptions for that situation, not findings about a whole group, and the number of questions on each topic follows what those people would ask, not how much the topic matters to the public.

Each study names the assistants it covers. We ask them through their ordinary consumer apps, the way a person would, rather than through developer access. Each question is a new conversation, so we capture the first answer a person would see.

Where a question depends on place, we can ask it from more than one location. Every assistant gets identical wording.

NZBrands.ai runs its studies on Onsomble, the platform that asks AI assistants questions at scale and records what they say. For each study we set out in Onsomble the subject, the questions and the people they are written as, the businesses or parties to track, the assistants, and the locations. A study about a subject, such as an election, has no focal entity, so everything tracked is treated alike.

Onsomble saves every answer in full and records the claims each answer makes, what each claim is about, and the sources behind it. We download the complete record after each run and build every published figure from it ourselves, with scripts that can rebuild the tables from the saved files.

Assistants are compared only on the questions every one of them answered, so a missing answer never changes what is being compared. Locations are compared only on identical questions. Counts about people or topics describe the study's coverage, not the public.

When an answer is missing, we say so and leave the gap. We never fill it with an estimate.

Each run is an edition, dated by the day the answers were collected. Every figure belongs to an edition, and the page says which one. We keep the questions the same between editions so that changes in the answers can be seen. If we have to change a question, we say so on the method page.

A study records what the assistants said on the day. AI answers vary, so the same question asked again can produce a different answer; one run shows what a person could have seen, not what they will always see. The people in a study are designed situations, not a sample of the population.

A study shows what was said and lets anyone check it. It does not show what anyone did because of it, or why an assistant answered as it did. Each study's method page sets out its own limits.

NZBrands.ai is a research publication from Onsomble. Businesses and parties cannot pay to be included, excluded or described differently. If we find an error in a figure or a claim, we correct it and note what changed and when.