How the NZ Google AI Overviews Study 2026 was built
Where the 1,000 searches came from, how we collected the results, what counted as an AI Overview, how we checked our own numbers and what the study cannot tell you. Read this before comparing our figures with any other AI Overview study.
What we set out to measure
How often, and how prominently, do Google AI Overviews appear when New Zealanders make commercially important searches, and does that change by industry or by what the searcher is trying to do?
The study is a snapshot of Google results on one day. It measures whether an AI Overview appeared and where it sat on the page. It does not measure clicks, lost traffic, revenue or how people behaved afterwards.
How we built the sample
We wrote 1,000 searches that a real customer or business buyer in New Zealand might type into Google, spread evenly across 20 industries and five kinds of search intent:
- 20 industries, 50 searches each;
- five search intents, 200 searches each;
- 10 searches in every industry and intent combination.
Giving every category the same number of searches is what lets us compare them fairly. It also means the overall figure is a comparison figure, not a measure of how New Zealanders actually distribute their searching. High-volume and niche searches count equally.
The industries
Banking and finance; insurance; property and real estate; legal services; health and dental; home services and trades; construction and renovation; travel and tourism; automotive; retail and consumer products; energy and utilities; telecommunications; business software and cybersecurity; professional and business services; education and training; recruitment and HR; agriculture and rural services; beauty and wellness; care services; and logistics, removals and storage.
The list covers high-consideration household and business purchases, regulated and unregulated services, products, utilities, local trades and nationally bought services. They are practical groupings, not an official industry classification.
The five search intents
| Intent | What the searcher wants | Example |
|---|---|---|
| Advice / informational | An explanation, a rule, a process or guidance on what suits them | "how does income protection insurance work nz" |
| Comparison / research | To weigh up options, providers or products | "compare power companies nz" |
| Cost / pricing | A price, fee, rate or cost range | "how much does a building consent cost nz" |
| Transaction / action | To buy, book, apply, switch, enrol or request something | "book a rental car auckland airport" |
| Local provider | A provider in a named New Zealand city | "plumber wellington" |
Each search was assigned an intent when it was written, before we saw any results, and none were reclassified afterwards.
Locking the list before looking
The full list of 1,000 searches was fixed before collection began, and a digital fingerprint of the list was recorded so anyone can confirm it was not altered. We started with 500, decided to double the sample while collection was running, and wrote the second 500 before looking at any results. Nothing was added, removed or reworded because of what we saw. An earlier 39-search pilot was kept out of every figure on the results page.
Where the results came from
Google results were collected on 8 September 2026 through DataForSEO, an independent provider of search-results and keyword data used across the search industry. Every search was run as a Google New Zealand, English-language, desktop search, with Google's AI Overview allowed to load fully before the page was recorded. A handful of searches hit a temporary Google error and were simply re-run with the identical wording.
Google describes AI Overviews as summaries that help people grasp a topic and explore supporting links, shown only when its systems judge them useful. See Google Search Central.
Tooling: DataForSEO and an OpenAI research model
We designed and directed the study. An OpenAI research model executed it under a fixed protocol: writing and locking the search list, running the collection, computing the statistics and performing the quality checks. Every rule was set before collection, every response from the data provider is retained in the audit files, and no other data source was used.
What counted as an AI Overview
A search counted as positive only when a genuine AI Overview appeared as its own element on the results page. AI-generated text tucked inside a "People also ask" box or another feature did not count, so if anything our figures are conservative.
Position on the page
For every AI Overview we recorded where it sat in the order of elements on the page: first, second or third, or lower. We also recorded whether it appeared above the first ordinary organic result. Position is about order, not screen space. It tells you the AI Overview came before the organic listings; it does not measure how many pixels it occupied.
What each AI Overview linked to
For every AI Overview we kept its length, the number of links it cited and the domains those links pointed to. The "who gets cited" section on the results page is a descriptive look at that data rather than one of the planned tests. The full generated text is not included in the downloads.
Search volume
After the list was locked, we looked up Google Ads monthly search volume for all 1,000 searches. Google only reports volume for common phrases, so a usable figure came back for 275 of them. We did not guess or fill in the rest.
The headline results treat every search equally, because that is what the balanced design is for. As a check, we also weighted the 275 searches with volume by that volume. The result, 27.0%, is reported on the results page as a sensitivity check. It should not be read as a national figure, because volume was missing for nearly three quarters of the sample and a few broad local terms dominate the rest.
How we checked our own numbers
- Every percentage comes with a margin of error. The overall figure of 73.9% has a margin of under three percentage points either way. Industry figures, based on 50 searches each, have margins of around 12 points, which is why we treat the industry table as a broad pattern rather than a ranking. The 100 industry and intent cells hold only 10 searches each and are shown for the pattern only.
- Intent versus industry. Standard statistical tests confirmed that search intent had a large effect on whether an AI Overview appeared and industry a small one. The first 500 searches and the second 500 gave results within a few points of each other. The full test statistics are in the downloadable workbook for anyone who wants them.
- Structural checks. Exactly 1,000 unique searches, 20 industries, five intents, 10 in every cell, no overlap with the pilot, no search returning more than one AI Overview, and every result confirmed as a New Zealand, English-language response.
- A second run. Shortly after the main collection we re-ran 100 of the searches, one from every industry and intent combination, chosen by a rule that could not be influenced by the results. The second run agreed with the first on whether an AI Overview appeared 93% of the time and on its exact position 91% of the time. That is strong short-term stability, and a reminder that individual results pages do change.
An earlier planned re-check accidentally selected only local-provider searches because of how the list was ordered. We kept it as a supplementary check rather than discarding it; it agreed with the main run 90% of the time. The balanced 100-search re-run is the one we report.
What it cost
We set a spending ceiling of US$30 before starting. The entire study, including the main collection, the doubling of the sample, retries, the search-volume lookup and both re-checks, cost US$3.76 in data charges. The itemised ledger is in the downloads. We publish it deliberately: research like this is cheap to run and only expensive to run honestly.
What the study cannot tell you
- It is a balanced sample of commercial searches, not a survey of everything New Zealanders type into Google.
- It is a one-day, desktop-only snapshot. Google results vary by time, device, location and the tests Google runs on different users.
- Location was set to New Zealand as a whole. We did not simulate a searcher standing in each named city.
- Writing the searches and assigning intents involved judgement on our part.
- Industry figures rest on 50 searches each and cell figures on 10, so fine-grained rankings are uncertain.
- Search-volume data was available for only about a quarter of the sample.
- Position on the page is about order, not how much space the AI Overview took up.
- The study measures whether an AI Overview appeared and where, not whether it was accurate, whether people clicked, or what it did to anyone's business.
Run it yourself
Everything needed to repeat or audit the study is in the download section of the results page: the full search list, the row-level results, the re-run comparison, the cost ledger, the analysis tables and the workbook. All of it is free to reuse under CC BY 4.0 with a credit and a link.
Want this run on your own keywords?
The free AI Visibility Audit applies the same intent-by-intent approach to the searches that matter to your business, across Google AI Overviews, ChatGPT, Perplexity and Gemini. It is the same approach behind my AI search optimisation work.
Book Your Free Audit