---
title: "How we ran this study"
publisher: "Onsomble Limited"
publication: "NZBrands.ai"
edition: 2
collected_on: "2026-10-12"
html_url: "https://nzbrands.ai/research/nz-election-2026/editions/2/method/"
---

# How we ran this study

Edition 2 · answers collected 12 October 2026.

[Read on NZBrands.ai](https://nzbrands.ai/research/nz-election-2026/editions/2/method/)

The same 174 questions were asked again on 12 October 2026. This methodology explains the collection, the comparisons with 5 October, and what each measure means.

## Study figures

| Measure | Value |
| --- | --- |
| questions | 174 |
| voter situations | 32 |
| AI assistants | 4 |
| answers saved | 793 |
| collected | 12 Oct 2026 |

## Collection and pairing

Edition 1 was collected on 5 October and Edition 2 on 12 October 2026, using unchanged configuration version 1. The four consumer interfaces were ChatGPT, Gemini, Google AI Mode and Google AI Overviews. The study collected answers, rather than following voters or measuring their behaviour.

The design contains 820 planned cells: 205 question-and-location combinations on each of four platforms. Of 174 questions, 31 regional questions also run in the relevant location, giving nine location settings besides the national setting. A temporal pair has the same saved question, platform and region, with exact wording verified. The archives have 815 and 793 answers; the temporal panel has 793 per edition; the common four-platform panel has 178 cells and 712 answers per edition.

The refresh has 27 cells where no AI Overview was produced: 22 were captured earlier and five are missing in both editions. Absent Overviews are not coded as no party, no source or zero words. All 124 within-round national-and-location pairs are complete in both rounds.

## Explicit party-name presence

Reach counts a party once per answer, including brief mentions, source labels and follow-up offers. Question text and URL targets are excluded. Both editions use the same case-sensitive names and aliases. We reviewed 140 contextual matches that could otherwise give misleading results. Generic National wording and Blind Low Vision NZ are excluded. The 17-party roster remains fixed for comparison.

A party-name set contains all parties counted in an answer. A changed set means at least one party appears or disappears. The measure does not establish substantive disagreement, recommendation, depth of treatment or a political effect. Absence does not show hostility, lack of eligibility or no relevant policies.

## Source counting

Literal sources are URLs preserved in response text. Associated sources are claim-linked references joined to the captured answer. Domains are normalised and counted once per answer; repeated URLs and claims do not inflate reach. Tracking parameters are removed. The measures remain separate because associations need not appear as visible URLs.

A changed domain set means membership changes, not that all material is new. The source shelf lists the ten websites associated with most answers in the selected edition. Deltas compare each listed website’s count on the same panel in the earlier edition. Each topic uses its own paired denominator. Source prominence does not estimate causal influence.

## Guidance-panel coding

The eight questions are PF017, PF034, PF044, PF056, PF072, PF089, PF104 and PF142. All 31 matched refreshed answers were read in full and compared with the earlier coding. Four behaviours overlap: explicit choice boundary, conditional party or bloc alignment, prioritised shortlist and elicitation of further voter information. An unconditional single-party vote instruction is separate. A boundary requires an explicit decline to choose or rank a winner, rather than a generic “depends” qualification.

The original panel had 32 answers. The longitudinal panel removes the absent refreshed PF142 Overview from the baseline coding, leaving seven Overview answers per edition. The codes are answer-level judgements. The purposive panel cannot support a platform-wide endorsement rate.

## Comparative reading and quotation checks

Every dossier reads the primary national-setting Gemini question in both editions, giving 64 full readings across 32 situations. Aggregate measurements cover all matched questions and locations assigned to the situation. All 64 dossier excerpts match preserved text. Selected pairs on other platforms support the guidance and factual chapters. The report does not claim semantic or factual review of every paragraph in all 1,608 answers.

## Public verification

Fifteen propositions were checked on 13 October against primary sources. The register distinguishes corrections, persistence, new misleading claims, omitted detail, errors not repeated and supported continuity or updates. Each judgement concerns the identified proposition; a supported date does not validate capacity projections in the same answer.

The 9 October Health New Zealand release was retrieved through primary-site indexed content after direct retrieval failed. Official regional pages corroborate the release date. Environment Canterbury's current reform section supplies its transition position despite older footer metadata. Retrieval limits are recorded.

## Interpretation limits

Authored situations are not a representative population sample. Consumer answers can vary across reruns, retrieval conditions, interface versions and timing. One collection per edition records snapshots rather than isolating causes. The report makes no statistical-significance claim, infers no voter persuasion and does not convert presence into support. Verified developments are distinguished from altered AI descriptions.

A design change should be recorded as a new catalogue version. Preserved gaps prevent missing coverage from becoming an apparent substantive shift.

[Evidence guide (Markdown)](https://nzbrands.ai/research/nz-election-2026/editions/2/evidence.md)

## Publication and corrections

Published and funded by Onsomble Limited. NZBrands.ai takes no view on any party or policy.

[Independence](https://nzbrands.ai/about/independence/) · [Corrections](https://nzbrands.ai/about/corrections/) · [Contact](https://nzbrands.ai/about/contact/)