## A — Methodology and glossary ### Collection and coverage The completed scan collected 1,094 of 1,156 planned answer cells, or 94.64%. All collected answers completed analysis. There were 62 collection failures, no duplicate or unexpected exported answer cells and no empty exported answers. The archive retains the missing-cell register. Failure causes for individual cells are not established by the exported aggregate status. | Platform | Planned cells | Collected | Missing | | --- | --- | --- | --- | | ChatGPT | 289 | 287 | 2 | | Gemini | 289 | 289 | 0 | | AI Overviews | 289 | 253 | 36 | | AI Mode | 289 | 265 | 24 | | Setting | Planned cells | Collected | Missing | | --- | --- | --- | --- | | NZ setting | 772 | 725 | 47 | | Auckland | 96 | 92 | 4 | | Wellington | 96 | 93 | 3 | | Christchurch | 96 | 93 | 3 | | Dunedin | 96 | 91 | 5 | ### Three question groups in the national setting The national setting collected 360 baseline recommendation answers, 110 explicit-priority answers and 255 journey-guidance answers, totalling 725. The cities collected 369 controlled recommendation answers. Those groups sum to the 1,094 collected answers. The 193 exact questions are preserved with their approved metadata. Each baseline persona/output plans sixteen national answers: four wordings across four platforms. The regional plan supplies the same twenty-four controlled prompts to four platforms in each city. The additional thirty priority questions and sixty-seven guidance questions run at the national setting. ### Recommendation coding and formula For lender L in a reporting group, **recommended answers** is the number of distinct collected result IDs whose reviewed answer positively proposes L for the requested task. **Collected answers** is the number of collected result IDs in that group, including those without a qualifying named recommendation. The optional inclusion rate is recommended answers divided by collected answers. A lender counts once per answer. Different lenders can count in the same answer. The measure is not exclusive share of recommendations. Conditional alternatives count when positively proposed. A citation alone, incidental name, contact directory, negative example or example of an unspecified current lender does not count. Adviser/distribution roles and expressly personal-loan alternatives are preserved separately. For support, a named benefit must relate to the existing mortgage: service, refixing, repayment changes, cash-flow management or genuine restructuring. Underwriting-only benefits are excluded. Explicit refinancing to resolve an existing repayment/restructuring problem may count as a conditional support route. List order is not treated as preference unless the answer explicitly makes a top recommendation. ### Comparisons and denominators | Comparison | Included evidence | Equal comparison unit | | --- | --- | --- | | National personas | 181 choice and 179 support answers | Actual collected answers per persona/output, 13–16 of 16 planned | | National platforms | 292 answers: 38 choice and 35 support questions × four platforms | Same question IDs on all four platforms | | Four cities | 340 answers: 43 choice and 42 support groups × four cities | Same question/platform groups in every city | | Platform within each city | Complete four-city groups split by platform | Same city questions within a platform; denominators differ between platforms | | Explicit priorities | 110 of 120 national answers | Separate case-specific matched priority prompts; one illustrative ChatGPT contrast discussed | | Journey guidance | 255 of 268 national answers | Question cohorts classified by primary stage; qualitative selected-answer review | No city repetitions or priority variants enter the pooled national persona rankings. All 369 city recommendation answers were coded, including the twenty-nine answers excluded from the equal-city comparison because their group lacked another city result. All 360 national baseline recommendation answers remain in the ledger, including questions excluded from the all-four-platform comparison. ### Review coverage and interpretation Every collected national and city baseline recommendation answer—729 records—was reviewed and coded for the requested recommendation role. The narrative chapters also use selected journey and priority answer passages; they do not claim exhaustive qualitative review of all 365 national priority/guidance answers or verification of every product assertion. The persona cases are purposive coverage, not a representative consumer sample. They have no population weights. The primary platform comparison equalises questions between platforms but does not create equal numbers of surviving questions per persona. The city comparison equalises the actual question/platform mix between cities. Each answer cell is one observation in this baseline. Matched settings make the observed comparisons fair; repeat collection is needed to assess persistence. Recommendation frequency does not measure audience reach, applications, market share, actual customer-service quality or purchase outcomes. ### Glossary | Term | Meaning in this report | | --- | --- | | Persona | Authored household/borrower circumstances relevant to the question | | Situation | The trigger, prior steps and current facts at a decision point | | Journey stage | The immediate task in the lending process or existing-loan review | | Intent | What help the question requests, such as understanding, comparing or resolving | | Buying factor | A criterion relevant to making progress, such as total cost or flexibility | | Controlled wording | A repeated request structure with the relevant circumstances supplied | | Priority variant | A matched question adding an explicit cost, flexibility or help priority | | NZ setting | The national collection setting, without a city in the prompt text | | Matched platform questions | Questions with a collected answer on every selected platform | | Matched city groups | Question/platform combinations with an answer in every city | | Recommendation inclusion | A favourable named proposal for the requested task | | Conditional alternative | A positively proposed option if a stated condition applies | | Incumbent | The lender already holding the mortgage | | Support drift | An answer that returns wholly or partly to acquisition despite an existing-customer setup | | Adviser route | A recommendation of help arranging or reviewing lending, distinct from funding | | Source association | A citation/reference record archived by the scan, without an asserted causal effect | ### Reproducible records National answer coding; city answer coding; national evidence and arithmetic validation; matched comparison validation; national aggregation script; platform/city aggregation script. ---