Short answer
AI market research works when the AI organises, calculates and summarises from sources you supply or open, and fails when you let it supply the numbers. For a New Zealand market question, start from official data such as Stats NZ business demography, have the AI show its arithmetic, check every competitor it names, and tag each claim as found, calculated, assumed or unverified.
- Never take a market size from the model’s memory; give it the source table.
- Tag every claim: found, calculated, assumed or unverified.
- Stats NZ counts enterprises and business locations differently; pick the one that matches your pricing.
- Open every competitor link the AI suggests; some will be overseas, closed or invented.
- Business demography figures used here are at February 2025; Stats NZ updates them on 30 October 2026.
- Goal
- Answer a market question with AI while keeping sourced numbers, your own calculations, assumptions and AI guesses clearly apart.
- Who it is for
- Owners, marketers, business development staff and analysts in New Zealand businesses testing an idea or a new segment.
- Time
- 30 minutes.
- You need
- ChatGPT, Claude, Copilot or Gemini, ideally with web search switched on, and the Stats NZ extract below.
- You will produce
- A segment size with the arithmetic shown, a checked competitor list, and a found-versus-inferred table.
What is AI good and bad at in market research?
AI is good at the slow, structural parts of market research: turning a vague question into sub-questions, summarising sources you give it, grouping customer comments into themes, listing competitors to check, and drafting interview questions. It is bad at supplying facts. Ask a model “how many cafés are there in New Zealand” and you will get a confident number with no reliable origin.
The rule for this lesson: the AI may find, organise and calculate, but every number must trace to a source you opened.
How do you separate found facts from guesses?
Tag every claim in your research with one of four labels. This one habit prevents most market research errors made with AI.
| Label | Meaning | Example |
|---|---|---|
| Found | Taken from a source you opened, with the reference date | 26,070 accommodation and food services enterprises at February 2025 (Stats NZ) |
| Calculated | Derived from found numbers, arithmetic shown | 3,156 + 3,018 + 1,257 = 7,431 enterprises with 6 to 49 employees |
| Assumed | Your stated assumption, to be tested | One in three of those would consider switching rostering tools |
| Unverified | Suggested by the AI, not yet checked | “Competitor X has 2,000 NZ customers” |
Anything left in “Unverified” when you finish stays out of the report.
How do you use AI on customer comments?
Customer comments are where AI saves the most time, because the source sits right in front of it. Paste 30 to 100 comments from reviews, survey answers or sales call notes, with names and contact details removed, and ask for themes with counts and two exact quotes per theme. Then search the raw comments yourself for two of the themes to confirm the counts.
Two failures to watch for: a “theme” built from a single loud comment, and quotes the model has quietly tidied up. Ask for exact wording and spot-check it. If the comments include customer details, read our lesson on protecting customer information when using AI first.
Exercise: size a niche from an official table
The business below is fictional. The numbers are real rows from Stats NZ’s business demography statistics at February 2025, Table 1.
Illustrative research brief. Fictional business.
Business: Kōkako Rosters, a Nelson software start-up
Product: rostering and food-safety checklists for hospitality
Target: NZ hospitality businesses with 6 to 49 employees
Pricing: per site (location), monthly subscription
Question: how many potential customer businesses and sites exist?
Source extract: Stats NZ, NZ business demography statistics,
At February 2025, Table 1. Industry H, Accommodation & food services.
Employee count size group: 0 | 1-5 | 6-9 | 10-19 | 20-49 | 50-99 | 100+ | Total
Enterprises: 11370 | 6894 | 3156 | 3018 | 1257 | 234 | 144 | 26070
Geographic units: 11649 | 7215 | 3411 | 3318 | 1632 | 309 | 102 | 27636
- Run the sizing prompt. Paste the brief and the worked example prompt below into your AI tool.
- Check the arithmetic. Add the three size groups yourself. Enterprises should total 7,431 and geographic units 8,361. Correct the tool if it differs.
- Read the caveats. Confirm the tool noticed that industry H includes accommodation as well as cafés and restaurants, and that the figures are from February 2025.
- Ask for competitors. With web search on, ask: “List rostering or food-safety software sold to New Zealand hospitality businesses, with the URL of each company’s own pricing page.”
- Open every link. Strike out any company that is overseas-only, closed, or whose page does not exist. Note the date beside any price you record.
- Fill in the four-label table for every claim you plan to use, and write a two-sentence answer to the research question.
Worked example: the sizing prompt
You are helping with New Zealand market research. Use ONLY the
numbers in the source extract below. Do not add statistics from
memory or elsewhere.
1. Calculate the number of enterprises and geographic units with
6 to 49 employees. Show the arithmetic.
2. Explain which of the two figures better matches per-site pricing.
3. List caveats about using this industry category for the target
market, including the reference date.
4. List assumptions we would need to test to get from this number
to likely customers, labelled ASSUMPTION.
[paste brief and source extract]
Example output (illustrative and abridged; your tool’s wording will differ):
1. Enterprises, 6-49 employees: 3,156 + 3,018 + 1,257 = 7,431.
Geographic units, 6-49 employees: 3,411 + 3,318 + 1,632 = 8,361.
2. Pricing is per site, so geographic units (business locations)
fit better. A cafe group with three outlets is one enterprise
but three geographic units.
3. Caveats: industry H includes hotels, motels and other
accommodation as well as cafes and restaurants, so 8,361 is an
upper bound. Figures are at February 2025.
4. ASSUMPTION: share of sites that are food service rather than
accommodation. ASSUMPTION: share without rostering software.
That output is useful because it refuses to go further than the data. The next numbers come from interviews and competitor research, not the model.
Where this goes wrong
- Invented statistics. The most common failure. A precise-looking percentage with no link is unverified, however confident it sounds. Our lesson on checking AI answers against sources covers the method.
- Overseas data passed off as local. Australian or US market sizes appear in New Zealand answers without comment.
- Stale figures. Annual releases are refreshed. Stats NZ’s release calendar lists business demography at February 2026 for 30 October 2026, so re-run your sizing after that date.
- The agreeable assistant. Models tend to support your hypothesis. Ask for the strongest case against it.
- Unchecked research agents. Agents that browse and compile reports still need evaluation. GTM-Bench, a 72-task benchmark for agents doing go-to-market research, was co-authored by Levy Street team members; our write-up on AI agents for go-to-market research explains what that work tests.
Check your work
- I recalculated the 6 to 49 employee segment by hand and got the same totals.
- I can explain when to use enterprises and when to use geographic units.
- Every competitor on my list has a working link to the company’s own site.
- Every number in my notes carries a found, calculated, assumed or unverified label.
- Nothing labelled unverified is in my final answer.
- I have noted the reference date of every official figure.
Where to go next
Next in the pathway: creating a source-checked content draft, which applies the same labelling discipline to writing. For researching one organisation rather than a market, see researching a customer account, and for working through a full data extract, analysing a spreadsheet with AI. When the sizing feeds an investment decision, put it into the AI business case template. If you want research like this automated and evaluated properly, our engineering guide to AI agent evaluations and our AI consulting team can help.
Sources
- New Zealand business demography statistics: At February 2025, Stats NZ, 30 October 2025
- New Zealand business demography statistics at February 2025: download data (Table 1), Stats NZ, 30 October 2025
- Release calendar, Stats NZ, Accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.