How we ran this study

We wanted to know what ChatGPT, Gemini, Google AI Mode and Google AI Overviews say when a voter asks them about the 2026 New Zealand general election: how they explain the election, which parties they bring up, how they compare policies, where their information comes from, and whether what they say is right.

This is the fourth edition. The questions, voters, assistants and settings are the same as on 5, 12 and 16 October, so differences between editions are differences in the answers, not in what we asked.

The study is neutral. No party is the focus, every registered party is treated the same way, and we take no view on any party or policy. We report what the assistants said.

Each question is written as a particular person would ask it. We described 23 households by their circumstances: their housing, work, income, caring responsibilities and access to services. We added nine people in particular locations, from a West Auckland renter to a Queenstown hospitality worker.

Each voter's priorities are our design assumptions for that situation. They are not findings about everyone in that group.

Every question was asked as a standalone conversation, so we captured the first answer a voter would get. The four assistants received identical wording.

The 31 local questions were asked twice: once with no location, and once from the voter's own location. That gives 205 question-and-location combinations and 820 planned answers per edition.

NZBrands.ai runs its studies on Onsomble, the platform that asks AI assistants questions at scale and records what they say. The study is set up as a topic with no focal entity, so no party is singled out, with every party on the Electoral Commission's register tracked. For this edition Onsomble re-ran the same saved setup: each question through the assistants' ordinary consumer apps, every answer saved in full, and for each answer the claims it made, which parties each claim was about, and the websites the assistant drew on.

We downloaded the complete record after the run and built every figure on this site from it. The script that does so is in the site's code, so the tables can be rebuilt from the saved files without running the study again.

Every ChatGPT and Gemini answer was saved. Google produced no AI Overview for 36 question-and-location combinations, 32 asked nationally and four from a location, and no AI Mode answer for eight, six asked nationally and two from a location. Each missing answer was requested again on its own and nothing came back; the record does not say why. Onsomble's run status counts one more answer than could be retrieved, and that answer is in no figure. A missing answer is left missing; it never counts as an answer that named no party.

116 national-and-location pairs came back complete in all four editions, and the location comparison uses those.

Every change compares the same question, assistant and location on 16 and 19 October. This edition compares the 776 answers that came back in all four editions, a smaller set than Edition 3 used, so its counts for 16 October can differ from those on the Edition 3 page. Every count covers the seven parties: 485 answers changed their party line-up.

Assistants are compared on the 163 question-and-location combinations where all four answered in every edition, 652 answers per edition. A missing answer never changes which questions are compared. Whole-edition counts, such as how many answers were saved, use all 776.

Locations are compared only within the 116 pairs where the question was identical. We never rank locations against each other on different questions.

The eight voting-guidance questions were read in full for all 31 answers and classified by hand, with the same definitions as Edition 3. For each of the 32 voters we read the Gemini answer to their main question on both dates and wrote up what changed.

We selected 15 propositions worth checking: recurring consequential claims and new issues found while reading. Thirteen of them appear as statements in this edition's answers, and those thirteen are listed in the checks: seven are supported by the source, three are contradicted by it, and three leave out a qualification the source makes. The other two are not stated in this edition's answers and are not shown as checks. Health New Zealand's pages refused automated access, so two health checks rely on the text of those pages as shown in search results.

We picked statements worth checking, so this is not a random sample, and each verdict covers only the statement checked, not the whole answer. A statement that is right today may still be out of date in an older answer.

Each question was asked once per edition. AI assistants vary their answers, so part of any change between editions is that variation rather than a lasting shift. Four editions show what changed; they do not show why. News events, what the assistants retrieved and changes to the assistants themselves all remain possible explanations.

The voters are designed situations, not a sample of New Zealanders. The study shows what the assistants said and lets anyone check it. It does not show whether anyone changed their vote.

We read all 31 voting-guidance answers and each voter's main question in full, and checked 13 statements. We have not read all 776 answers in full or checked every policy claim.

The archive holds all 776 answers, the 174 questions with their tags, the claims and sources Onsomble recorded, the guidance coding, the checks, the location pairs and the edition-to-edition comparisons, with file fingerprints so anyone can check the evidence has not changed since collection. Editions 1 to 3 keep their own archives and pages.