How We Surveyed 500 Australian Brands for AI — and Why It's a Floor
Good numbers deserve an honest account of how they were made. This survey involved two hard problems most "X% of businesses use AI" headlines quietly skip: you can't actually list every Australian business, and you can't cleanly detect an AI bot from the outside. Here's how we handled both, and exactly what the results can and can't claim.
Problem 1: there is no list of "all Australian businesses"
Australia has around 2.5 million actively trading businesses, and no register maps them to their websites. So "all Australian businesses" isn't a measurable population. We scoped to a defined, defensible universe instead: ~500 everyday-encounter brands — the banks, supermarkets, telcos, insurers, retailers, energy retailers, airlines, super funds and services most Australians actually deal with — across 21 sectors. The findings describe that population, not the whole economy.
Problem 2: detecting a bot from the outside is hard
Every chat widget loads vendor code, and that code is a fingerprint. We match each site against a signature table of chat and conversational-AI vendors, plus named-bot keywords, and sort each brand into a bucket:
| Bucket | Meaning | Counts as AI? |
|---|---|---|
| Strong-AI | Conversational-AI vendor or named/AI bot in the page | Yes |
| Has-chat (ambiguous) | A chat vendor that could be AI or human | Unconfirmed |
| Human live-chat | A human-only live-chat vendor | No |
| None / unreachable | No detectable widget | No (but may be hidden) |
Why we render every page
Our first pass read only the raw HTML — and missed a lot. Modern brand sites inject the chat widget with JavaScript after the page loads, so it never appears in the static source. In our pilot, NAB, Westpac and Woolworths all returned zero signal in raw HTML despite running real assistants. So we switched to a real-browser crawl (Playwright/Chromium) that loads each homepage and its support pages, waits for widgets to inject, and reads the fully rendered DOM, iframes and scripts.
Rendering helps — but it still doesn't catch everything, which is the whole point of the next section.
Why the numbers are a floor, not the truth
A crawl can only see a bot that loads on a public page. Three big categories stay invisible no matter how well you render:
- Behind a login. Macquarie's assistant Q lives inside the banking app.
- Behind a click, deep in support. NAB's assistant only appears after you open a messaging page and start a chat.
- Not on the web at all. Woolworths' Olive began life as a phone voicebot.
Every one of these failures undercounts — never the reverse. So we report our automated figure as a high-precision floor: read it as "at least this many."
Calibration: the hand-verified pilot
To turn the floor into an estimate, we hand-verified a 25-brand sample of the biggest names against public reporting and direct checks. That gives us a recall factor for the automated crawl and a true-rate anchor for the top of the market.
What this can and can't claim
The confirmed-AI floor is a lower bound. "Any chat widget" is a loose upper bound on AI (much of it is human or marketing chat). The ~28% likely-AI figure — floor plus conversational-AI platforms — is the defensible middle. Sector rankings are robust because detection precision is high; absolute levels are conservative because recall isn't perfect.
Everything here is reproducible: a curated brand universe, a render crawl with a published signature table, a verification pass, and a report. The honest summary is that we can tell you confidently which sectors lead and lag, and give a well-bounded estimate of how many Australian brands run an AI assistant — while being upfront that the true number sits above the floor we can prove.
Back to the findings: AI adoption across Australian business.