Australia got two counts of its AI users on the same day, and on Google Gemini they are four million apart.

On 2 June 2026, Roy Morgan published a finding that 13.6 million Australians aged 14 and over use AI tools in an average four weeks. The same day, Telsyte published a study putting the Australian AI user base at 17.4 million, on a population base that starts two years older.

Neither is wrong. Working out why they differ is the most useful hour anyone buying AI visibility in this market can spend, because a vendor quoting one of these numbers at you is quoting a definition as much as a measurement.

What each study found

Roy Morgan's finding comes from Single Source for January to March 2026, a sample of 14,646 Australians aged 14 and over. It reports 13.6 million people, 58 per cent of that population, using an AI tool in an average four weeks. ChatGPT leads on 10.5 million and 45 per cent, which Roy Morgan describes as more than twice as popular as the next tool. Google Gemini is second on 5 million and 21 per cent, then an estimated 4 million on Microsoft Copilot, an estimated 1.4 million on Canva Magic Studio and an estimated 777,000 on Claude.

Telsyte surveyed 2,023 Australians aged 16 and over, online, across April and May 2026. It puts 17.4 million Australians, 77 per cent of the 16 and over population, using AI, with the user base up by more than 6 million in the year to May 2026. Its platform table runs ChatGPT 13.8 million, Google Gemini 9.1 million, Meta AI 5.6 million and Microsoft Copilot 5.4 million, then Apple Intelligence, Samsung Galaxy AI, Claude, Canva AI, Grok and CapCut.

Gemini is the widest gap in proportional terms. Telsyte's figure is 82 per cent higher than Roy Morgan's for the same platform in the same country, measured six weeks apart.

The part that resolves most of it

Telsyte publishes a second figure that almost nobody quotes, and it is the one that matters. Of that 17.4 million user base, 14.3 million engage with AI at least monthly, and 5.2 million use it every day.

Roy Morgan's 13.6 million is a four-week measure. Telsyte's 14.3 million is a monthly-engagement measure. Compare those two rather than the headlines, and two studies that looked 3.8 million apart come within about 700,000 of each other, on different samples, different modes, and windows two months apart. At the national level the disagreement is mostly a definition of the word user.

That does not rescue the platform figures. Telsyte's per-platform numbers count its whole user base, so Gemini's 9.1 million and Roy Morgan's 5 million are not describing the same behaviour, and neither publisher gives a figure that lets anyone reconcile them. Second place stays unresolved.

A third source, and a third order of finish

Statcounter's Australian AI chatbot share for June 2026 has ChatGPT at 67.7 per cent, Microsoft Copilot second on 14.2 per cent, Google Gemini third on 7 per cent, Claude on 6.1 and Perplexity on 5.1. We read those figures in a post by the Melbourne agency Optimising rather than at Statcounter, so they are cited at one remove and flagged here as such.

Statcounter measures referral share rather than people, so it is a different unit again. Worth having anyway, because it puts a third ordering on the same market. Gemini is second on both surveys and third on traffic share, while Copilot is third and fourth on the surveys and second on traffic.

What none of them can tell you

Not one of these datasets says whether your business is mentioned when an Australian asks about your category. They count people and sessions. Brand presence inside an AI answer is a separate measurement problem, and the IAB conceded on 3 August that the industry has not standardised it.

Roy Morgan attaches a caveat to its Gemini figure that deserves more attention than it gets. The number "represents the active use of the tool, rather than the embedded use that comes with using Google Search." So 21 per cent counts people who opened Gemini deliberately. The AI summary that appears above a search nobody asked to have summarised is in none of these tables.

Survey mode matters too, and this next part is our reading rather than either publisher's. Telsyte's sample is 2,023 people recruited online. Roy Morgan's is 14,646 through a continuous survey. An online panel reaches people already comfortable online, which is the population most likely to have tried an AI tool.

What we think this means

All three sources agree on one thing: ChatGPT leads the Australian market and it is not close. Roy Morgan has it at more than double the next tool, Telsyte at 1.5 times, Statcounter at nearly five times on traffic share. First place is robust across three methods. Everything after first place depends on who measured it and how.

That is enough to act on and it is not enough to buy software on. If a client asks where to start, the answer is ChatGPT, because all three methods put it first by a distance and checking costs nothing: open it, ask what a customer would ask, and write down what comes back and the date you asked. When a vendor puts an Australian AI number in a deck, the question back is which base, which window, and what they counted as a user. On the evidence above those three answers move the same figure by four million people.

There is a version of this where the disagreement is a reason to wait. We do not read it that way. Both studies have Australian AI use as majority behaviour, and Telsyte has the user base growing by more than six million people in a year, so the direction is not in dispute and the precision is. A business that waits for the sources to converge is waiting on a measurement argument rather than on a customer.

Roy Morgan published a follow-up on 14 July that is the one we would put in front of a client. AI users are more likely to be employed full time, 51 per cent against just over 40 per cent of all Australians, with median personal income of $67,000 against $53,000 and median savings and investments of $181,000 against $149,000. Michele Levine's own description is "a valuable demographic for advertisers to reach". Anyone who has been told their buyers are too old or too conservative for any of this should read those three numbers again.

The practical reading is narrow. Ask your customers' questions in ChatGPT, record the date, and if you do not come up, that is a content problem before it is a measurement problem. We have written on what earns a citation in an AI answer and on the free tool Google now gives you for the search side of it. Neither piece had Australian numbers under it until now, and now it has two sets that do not agree.