What the research found
What IF Web ran three surveys to New Zealand consumer panels between August and September 2026, asking 300 people to type the exact words they would put into an AI tool. That produced more than 1,200 written prompts across three unrelated categories. Between 72% and 77% of respondents use AI tools at least weekly, and 28% to 29% use them every day. ChatGPT leads at 71% of AI users and Gemini sits second at 45%. The median prompt runs between 9 and 14 words, fewer than 10% are short keyword strings, and in one survey only 2% of people named a brand at all.
One of the three surveys was run for What IF Web. The other two were run for clients, and their categories are described in general terms here.
Why survey people instead of pulling prompt data from a tool?
Because the data does not exist. Outside Bing's grounding queries, no major AI platform publishes what people are prompting for. There is no Search Console for ChatGPT.
Every prompt list you have seen has been reverse-engineered from keyword volume, which tells you what people typed into a search box rather than what they say to a chatbot. Those are different behaviours, and the data below shows how different. So we started asking people directly, and this kind of research is now part of the What IF Web AI Search service for clients who want their tracked prompt set built on real language instead of a keyword tool's best guess.
How often do New Zealanders use AI tools?
Between 72% and 77% use them at least weekly, and 28% to 29% use them every day.
Two of the three surveys ran three weeks apart with completely unrelated audiences, and they landed within a few points of each other on both measures. When two independent samples agree that closely, the number is worth trusting.
The wider variation sits at the other end. In one survey 14% said they never use AI tools. In another it was 4%. Whether your buyers have picked this up depends heavily on who your buyers are, and a national average will hide that from you.
Which AI tools are New Zealanders using?
ChatGPT dominates at 71%, and Gemini is a much stronger second than most people assume.
Figures are across 182 respondents who use AI tools, from the two surveys that asked which tools they reach for. Respondents could tick more than one, so the column does not total 100%.
Gemini is doing a lot of quiet work at 45%. It came second in both surveys by a wide margin and gets a fraction of the attention ChatGPT does in New Zealand AI search conversations.
The Google AI Overviews figure of 33% is almost certainly understated, because people do not know they are using it. They typed a search into Google, read the answer at the top, and would not describe that as using an AI tool. Some of them may also be ticking Gemini for the same experience, since Google has not gone out of its way to keep the two separate in anyone's head. Treat that number as a floor rather than a measurement.
How long is a typical AI prompt?
The median prompt in our surveys ran between 9 and 14 words, depending on the category. Fewer than 10% were short keyword strings of four words or fewer.
That is a long way from the two or three words people habitually type into Google. Most wrote one full, plain sentence.
Worth being careful about what that measures. We captured opening prompts, and an opening prompt is the start of a conversation rather than the whole of it. In real use people ask, read the answer, then narrow it: which of those is cheapest, do any of them work in Christchurch, tell me more about the second one. The opening prompt decides which names appear on the screen. Everything after it decides which of those names survives, and we only measured the first move.
A meaningful minority front-load far more. In our own survey, 27% of prompts ran past 20 words, and those people volunteered things a search box never asked for: their budget, their suburb, how many people live in the house, what system they already have, what went wrong with the last supplier. Several asked the model what information it needed before it answered.
One detail did surprise us. Only around 10% of prompts ended in a question mark. Plenty were instructions rather than questions: "find me a company that", "help me find", "please suggest". Question-shaped headings are still the right instinct, they are just covering some of this behaviour rather than all of it.
Which words do people put into their prompts?
Location, price, and trust, with the mix shifting hard by category.
Local intent showed up in about 40% of prompts in the consumer product survey as a "near me", a city name, or "NZ". In our own professional services survey it was closer to 12%. Same country, same month, wildly different geographic instinct, because you buy groceries near you and you buy a website from wherever.
Price language runs through everything: cheapest, affordable, best value, reasonably priced. So does trust language, and it is blunter than you would expect from people typing to a machine. Reputable. Good reviews. Highly rated. One respondent asked for companies "with a good reputation in Customer reviews and not a scammers. Legitimate websites." (Punctuation theirs, sentiment universal.)
Where the decision involves choosing between options, the framing is near-identical every time. Pros and cons. Compare. Which is better. Which is cheaper. If your buyers face that kind of choice, a real comparison table on the page is one of the highest-value things you can build. Real tables extract at roughly 81% accuracy against 23% for the same information written as prose, which is a gap worth caring about.
How much should you trust these numbers?
Less than you would trust observed behaviour, and more than you would trust a hunch.
These are self-reported prompts typed into a survey. People almost certainly write more carefully in a survey box than in a chat window at 9pm. Each survey ran to 100 respondents, so treat everything here as directional rather than precise. We covered three categories, not thirty, and the size of the variation between those three is the argument against stretching it further.
What the data does support is a starting point. Run this research for your own category and you get the phrasing your buyers use, which beats guessing at prompts from a keyword tool.
This research was run by What IF Web, a Webflow Premium Partner studio in Christchurch, New Zealand, and New Zealand's first Otterly agency partner. We build websites and brands for businesses that are done settling, and we run AI search visibility work for clients across New Zealand.
Notes on the research
Three surveys were run through Pollfish to New Zealand consumer panels between 20 August and 1 September 2026, at 100 respondents each. One was run for What IF Web. The other two were run for clients, and their categories are described only in general terms. Open-ended questions asked respondents to type the exact wording they would use with an AI tool. Prompt length, modifier frequency, and brand-mention rates were calculated from the cleaned response sets, after removing respondents who reported never using AI tools.
