Twelve per cent sounds like a rounding error. It is not.
That is the share of Australians who now name an AI tool as their main way of finding information online, from Telsyte's Australian Artificial Intelligence Study 2026. Twelve months earlier the same study put it at 5%.
I have been doing search work in this country for more than twenty years. Numbers in this industry usually move the way glaciers move. A measure of consumer behaviour that more than doubles in a year is not a trend line, it is a different market.
And almost none of it appears in the reports most Australian businesses read every month.
Two Australian numbers that disagree
A caveat first, because you will see both of these quoted and they do not match.
Roy Morgan has 13.6 million Australians, or 58% of people aged 14 and over, using AI tools in an average four weeks in the March quarter of 2026, from a sample of 14,646. Telsyte has 17.4 million, or 77% of the population aged 16 and over, with the user base growing by more than six million between June 2025 and May 2026.
Both are credible. They are counting different things. Telsyte's own top ten includes Apple Intelligence at 3.9 million users and Samsung Galaxy AI at 3.1 million, which are features that arrive switched on inside a phone. Roy Morgan's list is dominated by tools a person deliberately opens: ChatGPT at 10.5 million, Gemini at 5 million, Copilot at 4 million.
If you sell something, the second measure deserves your attention. A person who opens ChatGPT to work out who to call is a person who did not open Google.
Every figure here is from the Telsyte Australian Artificial Intelligence Study 2026, an online survey of 2,023 representative Australians aged 16 and over fielded across April and May 2026.
Key figures from Telsyte's Australian Artificial Intelligence Study 2026: 12% of Australians name AI as their main way to find information online against 5% a year earlier, 17.4 million Australian AI users, 5.2 million daily users, 81% aware of AI summaries in search, and 27% who have had AI complete a task for them.
Why none of this shows up in your reporting
Here is the uncomfortable mechanic. An AI answer happens before the click, and your analytics only starts counting at the click.
Telsyte found 81% of Australians know AI summaries appear in search results, and half said they would often rely on the summary without clicking through to the source. That is not people failing to find your website. That is people getting what they needed and moving on.
The click data backs it up. Similarweb reported in June 2026 that 68% of Google searches now end without a click, rising to 80% on desktop. When an AI Overview appears, and it shows on more than 20% of queries, the probability of a click falls from roughly 15% to 8%. Google's AI Mode, which arrived for Australian users in October 2025, sends a visitor through on 1.6% to 2.5% of queries. Ordinary Google search sends one through on 17% to 19%.
Sit with that last pair for a second. On the surface Google is pushing hardest, the referral rate is about a tenth of what search used to give you.
There is a second layer underneath. A single AI Mode prompt does not run a single search. It fans the question out into a set of related searches, runs them in parallel, then writes one answer from whatever comes back. Analysts who have pulled the lid off this put the typical spread in the range of nine to eleven sub-queries for a complex question. None of those appear in Search Console. So you are not only missing the clicks, you are missing the questions.
When you open your analytics, see a handful of sessions from ChatGPT, and conclude AI is not worth worrying about yet, you have measured the exhaust rather than the engine.
The sequence behind an AI-first enquiry. Your analytics can only see the last step, and only if there is a last step.
A five step sequence showing how an AI-first enquiry forms: the buyer asks an assistant, the assistant runs searches nobody sees, reads mostly third-party pages, names two or three providers, and only then does the buyer make contact.
The agentic part is not a forecast
I have sat in meetings where AI agents doing the buying got filed under things to look at in a few years. In Australia it is already a quarter of the market.
Telsyte found 27% of Australian AI users have had AI complete a task on their behalf, including product comparisons and online shopping. Another 39% would let AI finish a task as long as they approve the final step, and among daily users that rises to 66%.
An assistant doing the comparison step is an assistant deciding the shortlist. You are either on it or you are not, and nobody sends you a notification either way.
The surface is being monetised too. Telsyte notes OpenAI is already testing advertising inside ChatGPT in Australia and the United States. Follow that sequence through. A paid position is arriving on a surface where most Australian businesses have never once checked whether they show up for free.
The trust gap is the actual opportunity
The same study that shows how fast Australians picked up AI also shows how little they trust it. Only 31% are comfortable with AI playing a larger role in managing their everyday activities. 62% feel technology is changing faster than they can keep up with. Just 35% trust technology companies to use their data responsibly.
On advertising, 70% would take a free ad-supported service over a paid ad-free one, yet only 27% are comfortable seeing ads inside AI-generated answers, and 52% think a sponsored AI answer is less trustworthy than an organic one.
I read that as a pricing signal. The unpaid mention is the one Australians believe. When an assistant names you because the web says you are worth naming, you have something an ad budget cannot buy quickly, and it is the slower of the two to build. Which is the argument for starting now rather than when it becomes obvious.
The part nobody wants to hear about fixing it
The instinct, once you find out an assistant is not naming you, is to go and write more pages. I understand the instinct. It is also mostly wrong.
A study published this year examined 167,551 citations sitting behind AI answers about 128 brands across 12 markets. It found 85.7% of those citations pointed at websites the brand does not own. Only 14.3% pointed at the brand's own site. The sources also clustered hard, with 80% of citations coming from roughly 18% of domains.
Read that as a work order and it says two things. Your own website is table stakes, and it is not where the model mostly reads you. And there is a short list of places that decide most answers in your category, so the job is finding that list rather than publishing into the void.
That is a bigger piece of work than one article can carry, so it is getting its own.
What I would measure instead
If AI referral sessions are a poor proxy for AI influence, then the work is to measure the answer itself. This is how we do it, and some of it was learned the hard way.
Ask the engines the questions your buyers actually ask, in their words, and record whether you get named. Not your brand name. The unbranded question, the one where you hold no advantage.
Keep recall and retrieval apart. A model answering from memory is telling you what it absorbed months ago. The same model with live web search switched on is telling you what the web says about you today. Those are two different facts about your business, and averaging them gives you a number that describes neither. We hold them as a pair and report them as a pair.
Write down which sources the engine cited. The verdict on its own is not much use. If a competitor gets named and you do not, the citation list usually tells you which third party made that call, and that is where the work is.
And treat a failed measurement as a failed measurement. If a run comes back with no sources at all, that is not a score of zero, it is not measured. Log it as a zero and you will spend a fortnight chasing a collapse that never happened.
None of this needs a big budget. It needs someone to actually look, on a schedule, and write down what they see.
The measurement set we keep for AI visibility. None of it needs a large budget, only someone looking on a schedule.
Checklist of six measurement practices for AI search visibility: probe engines with unbranded buyer questions, record who is named, keep training recall separate from live web search, capture cited sources, mark runs with no sources as not measured, and re-run on a fixed cadence.
Where this leaves you
Twelve per cent is not the number that should worry you. The doubling is.
If AI-first discovery went from 5% to 12% in a year, the useful question is not whether it reaches a quarter of the market. It is what you want the assistant to say about you when it does. That answer is being assembled right now out of the web as it currently stands, and you get no vote in it unless you have done the work.
The cheapest version of this is also the most uncomfortable. Open ChatGPT, open Gemini, and ask the question your best customer would ask. Then read the answer as a buyer rather than as the owner.
If your name is not in it, you have found your project for this quarter. If you want the longer explanation of how answer engines decide who to name, we set that out in our guide to SEO, AEO and GEO. If you would rather someone else did the looking, that is what an AI visibility audit is for.
Sources
- Telsyte, Australian Artificial Intelligence Study 2026, published 2 June 2026. Online survey of 2,023 representative Australians aged 16 and over, fielded across April and May 2026. Announcement
- Roy Morgan, AI tools usage, March quarter 2026. Sample of 14,646 Australians aged 14 and over. Findings
- Similarweb, zero-click and AI search referral data, June 2026. Analysis
- iPullRank, analysis of query fan-out in AI search. Research
- How Large Language Models Source Brand Reputation Across Languages and Markets, 2026. Analysis of 167,551 citations covering 128 brands across 12 markets. Paper
