How we researched this.
ClearLogic’s trade-industry research draws on publicly posted customer reviews from NoCowboys, a major New Zealand tradesperson-review platform. This page states exactly how we collected, scored, and filtered that data, so any figure we publish can be checked rather than taken on faith.
Publicly posted reviews, crawled politely.
We crawled NoCowboys respecting the site’s published robots.txt,
including its explicit allowance for AI agents and its 5-second crawl delay. We did not access anything
not publicly visible on the platform, and we did not create an account or interact with the site beyond
reading published pages.
The corpus covers three trades, national scope, no geographic pre-filtering:
| Trade | Businesses | Reviews |
|---|---|---|
| Electricians | 100 | 5,731 |
| Plumbers | 119 | 7,970 |
| Painters | 150 | 6,585 |
| Total | 369 | 20,286 |
This is a cross-sectional snapshot of each business’s full public review history as it stood at crawl time. It is not a rolling monitor and does not update automatically — if we republish an updated version later, we’ll date it separately from this one.
2006–2026, one known gap.
98.7% of reviews carry a parseable date, spanning 2006 to 2026. 45% of dated reviews fall in 2021 or later; 19% in 2024 or later. Electricians’ review dates are recorded in a text format our current extractor doesn’t parse, so year-by-year trend figures for electricians aren’t reported in our research until that’s fixed in a future re-crawl. We say so wherever it’s relevant rather than leaving a silent gap.
How reviews are scored, and how we coded them.
The scoring system
Each review on the platform carries four sub-scores — communication, quality, reliability, and value — plus an overall score, all on a 0–100 scale, self-reported by the reviewer at the time they left it. We didn’t create or adjust this scoring system; we report it as published.
Complaint & praise coding
We defined a negative review as one scoring below 50/100 on any dimension, and a positive review as one scoring 80/100 or higher overall. Within those groups, we coded the reason for the score against a fixed taxonomy — 5 complaint categories, 13 praise categories — using an LLM classifier (Claude Haiku 4.5, run in batches) rather than keyword matching, so two reviews describing the same problem in different words land in the same category. The taxonomy is fixed in advance of coding, not built from the results.
A statistic only gets published if it clears two thresholds.
- 01At least 30 reviews behind any single figure (a trade × region, trade × business-size, or similar breakdown), so a figure isn’t resting on a handful of opinions.
- 02At least 5 different businesses contributing to that figure — our own addition, on top of the review-count threshold, because a stat built from 2–3 businesses’ reviews can be traced back to roughly which businesses it describes even with no name attached. Breakdowns that don’t clear 5 businesses are dropped from what we publish, not estimated or smoothed over to fill the gap.
No names, ever.
No business name, owner name, staff name, or individual review appears in anything we publish. Every figure we report is an aggregate across many reviews and many businesses. Where we’ve used short illustrative language drawn from review themes, it’s written to describe a pattern our coding found — never quoted from, or close enough in wording to be traced back to, one specific review.
Read the numbers with these in mind.
- —This is public reviews, not a random sample of every job done. People with an unremarkable, fine experience are less likely to leave a review at all than people who were thrilled or angry, so the corpus skews toward strong opinions in both directions.
- —Correlation isn’t causation. Where our research notes a pattern between two things — for example, business size and average score — that’s an association in the data, not a tested explanation for why it happens.
- —Small figures on small bases are directional, not precise. Where a category is only supported by a small number of reviews, we report it as a rank or a pattern rather than a specific percentage.
Figures drawn from this corpus are treated internally as candidate findings until reviewed and approved for publication. This page will be updated if the underlying data is refreshed or re-crawled. Last updated: 7 August 2026.
