Methodology · NZ Election 2026 · Edition 5
How we ran this study
We repeated the election questions on 23 October 2026. This page explains what we asked, what came back, how we counted party names and how we compared the answers with 19 October.
- questions
- 174
- voter profiles
- 32
- AI assistants
- 4
- answers recorded
- 786
- collected
- 23 Oct 2026
01
What we set out to learn
What AI tells a New Zealander asking for help with the election, and what changed when we asked again.
In the general method:The purpose of a studyWe asked ChatGPT, Gemini, Google AI Mode and Google AI Overviews about the 2026 New Zealand general election. We looked at the parties they named, how they connected policies to voters' priorities, and selected factual statements.
This is Edition 5. We take no view on any party or policy, and no party pays to appear.
02
Who we asked as
The questions describe 23 household situations and nine people in particular locations.
In the general method:Writing the questionsThese are made-up situations designed for the study, rather than a sample of real voters. They cover housing, work, income, caring responsibilities and access to services.
The priorities written into a question belong to that situation. They are not assumptions about everyone with the same background.
| Question set | Questions | Purpose |
|---|---|---|
| Household situations | 122 | Practical concerns of the 23 households |
| Shared national questions | 21 | Election questions without a particular voter described |
| Regional situations | 31 | Local concerns of the nine regional voters |
| Total | 174 | 20 topics across 32 voter situations |
03
How the questions were asked
The same words, asked separately, through four consumer AI assistants.
In the general method:The assistants we coverEach question was asked as a standalone conversation. The regional questions were asked nationally and from the relevant location, giving 205 question-and-location combinations for each assistant and 820 planned answers.
The nine collection locations were Ashburton, Auckland, Dunedin, Hastings, New Plymouth, Queenstown, Tauranga, Wellington and Whangārei. The regional situations also name their local context within the question.
04
How we used Onsomble
Onsomble asked the questions and saved the complete answers for comparison.
In the general method:Working with OnsombleWe used Onsomble to repeat the study's saved question set. The wording and collection settings were checked before comparing the answers with the earlier collection.
We asked the consumer versions of these four services.
05
What came back
We recorded 786 of the 820 planned answers on 23 October 2026.
We could not obtain a complete answer for one ChatGPT request and 33 AI Overview requests. The individual reasons were not recorded, so we do not describe those gaps as refusals or as answers that named no party.
The collection finished on 23 October. No later replacement answers are included in this edition.
| AI assistant | Planned | Recorded |
|---|---|---|
| ChatGPT | 205 | 204 |
| Gemini | 205 | 205 |
| Google AI Mode | 205 | 205 |
| Google AI Overviews | 205 | 172 |
| Total | 820 | 786 |
06
How we count
The figures count party names in answers, rather than recommendations to vote for those parties.
In the general method:Counting rules- Answer
- The complete saved response to one question, from one assistant and one location.
- Party named
- A party counts once per answer when an unambiguous name or alias appears. Repeated names do not add to the count. We exclude web addresses, but include visible citation labels and follow-up text. Naming a party does not mean recommending it.
- Names with other meanings
- We check the context when a name can mean something else. National counts when it names the party; a reference to the National Library does not. The same principle applies to other party names.
- The seven displayed parties
- National, Labour, Greens, ACT, NZ First, Te Pāti Māori and The Opportunity Party. The evidence downloads retain all 17 tracked parties.
- Tone
- The coloured bars summarise Onsomble’s saved assessments of statements about each party: positive, neutral, negative or mixed. Only statements where the party is the subject contribute to its bar. Strongly positive and strongly negative assessments join the corresponding positive and negative groups. These are automated assessments of how a party is portrayed, not a measure of which party the assistant prefers.
- Sources
- A website counts once per answer when it is recorded as a citation. Other references are preserved in the downloads and excluded from these rankings. Cited websites are recorded for 784 of the 786 answers; two have an empty recorded list. A visible link in the answer is a separate observation. A recorded citation does not, by itself, prove that it supports every statement in the answer.
- Kinds of website
- The labels group websites as party, government, news, comparison, advocacy, business, social or other. Other includes websites we have not yet classified.
07
How we compared the answers
The main comparison uses answers to the same question, from the same assistant and location, on 19 and 23 October.
In the general method:Making comparisonsThere are 771 answers with an earlier counterpart: 204 from ChatGPT, 205 from Gemini, 197 from Google AI Mode and 165 from AI Overviews. Fifteen current answers are outside that comparison. Five earlier answers have no current counterpart.
Edition 4 contains 776 answers. This comparison uses the 771 for which we also recorded an answer on 23 October. The five without a later counterpart are left out of both sides of the comparison.
The source comparison also uses the same 771 answers on each date. Its counts come from the source lists saved with those answers. The original Edition 4 source figures remain on its own pages.
The four-assistant chart uses the same 159 question-and-location combinations for every assistant on both dates, giving 636 answers per date. The current location comparison uses 121 complete national-and-location pairs and does not claim a change since the earlier collection.
We used AI-assisted review of 31 current guidance answers and 32 main voter-case answers. Earlier recorded readings were reused after their text was checked, rather than all being read again. The extracts on each voter page open the original answer from the date shown.
08
What we checked
Nine selected checks cover eight claims from the 23 October answers.
We compared selected statements with official sources on 27 October. Four checks found an incorrect statement and five supported the statement checked. Two of the checks concern the same FamilyBoost claim in different answers.
The checks were selected because the claims mattered to the voter situations. They are not a random sample and do not estimate the accuracy of all 786 answers. Each verdict applies to the particular statement, rather than the whole answer.
09
What this can and cannot tell you
This is a dated record of answers, rather than a prediction of what every voter will hear.
In the general method:LimitationsAI assistants can answer the same question differently. One answer from each setting cannot show whether a change is lasting or whether location caused it. The study does not measure influence on votes or establish political bias.
All 786 answers were included in the counts of party names. Detailed interpretation covered 63 answers, not every answer in the collection. The AI-assisted readings were not independently assessed by a second reviewer.
The study does not verify every policy statement or establish anyone's personal entitlement. An earlier error disappearing from a later answer is not, by itself, a correction.
10
The records
Every saved question and answer stays available with the tables behind this edition.
The evidence pages include all 174 questions, the 786 complete answers, comparisons, recorded readings and selected checks. Downloads preserve the original research records separately from these reader summaries.
Each earlier edition retains its own findings, voter pages and original answers. Read an older answer in the context of its collection date.