Do AI assistants favour local suppliers inside Australia, and the contrast that explains why

Yes, measurably, in this market and on these two engines. Across 64 blind readings taken for Australia on 22 October 2026, every agency named in the answer prose was an Australian firm and no overseas supplier was named once. That is a clean result and its limits need stating immediately: two engines, one date, and the preference appears to track which pages are visibly about this market rather than where a company is registered.

The measurement, with its method

We put 33 batches of buying questions to ChatGPT and Gemini with Australia written into the text of every question. The market had to go in the prompt because the search interface we use accepts a country parameter, a language parameter and a location parameter, returns HTTP 200 for each, and ignores all three. The snapshot taken that evening holds 64 blind readings, 32 on each engine, with the brand batch excluded because a question naming us is a ceiling rather than a result.

What was named	Readings, of 64	Where it is based
Online Marketing Gurus	12	Australia
Rocket Agency	8	Australia
StudioHawk	8	Australia
Prosperity Media	8	Australia
Sparro	3	Australia
King Kong	3	Australia
Any overseas agency	0

The tools named were a different story and the contrast is the key to the whole thing. Semrush appeared in 14 readings, Ahrefs in 12, HubSpot in 11, Otterly in 7, Peec in 4, BrightLocal in 3, Profound in 2, and Scrunch, SE Ranking and Rankscale in 1 each. Not one of those is an Australian company.

What the contrast actually shows

The engines are not applying a nationality rule. They applied a local preference to the agency question and no local preference at all to the tool question, in the same readings, on the same day. The difference is in the question: an agency is a local service and a tool is not, so a local answer is the correct answer to one and not to the other.

That reframes the finding usefully. What the engines favour is not Australian ownership. It is a page that is obviously the right answer for an Australian buyer. A foreign agency with no Australian specific published work fails that test and a foreign tool with a global product does not need to pass it.

What the engines read to get there

Of 1,010 distinct hosts returned across the 64 readings, 435 carried an Australian domain. The commercial hosts returned most often were Australian specialist sites: localdigital.com.au in 9 readings, digitalnomadshq.com.au in 8, 3pdigital.com.au in 7, soupagency.com.au in 7, firstpage.com.au in 7, odindigital.com.au in 6, profitgeeks.com.au in 5, rocketagency.com.au in 5.

And above them, the state. Sixty six distinct Australian government hosts appeared 139 times across 25 of the 64 readings, with business.gov.au in 15, accc.gov.au in 13, acma.gov.au in 8, oaic.gov.au in 6, legislation.gov.au in 5 and state portals from Queensland, New South Wales, Victoria and others. That is the heaviest government signal in any of the seven country readings we took, and it is the structural reason a local answer looks local: the sources framing the question are Australian institutions.

What a foreign supplier should do with this

• Publish the Australian page, not a dropdown. Name the market, the currency, the service geography and the regulator where it applies. A global page with a country selector is quotable for nobody.
• Put an Australian reference on it. One, contactable, with permission. It converts the page from a claim into an answer.
• State the tax and currency position in a line. A price with no tax position reads as evasion in a market that checks.
• Answer the verification gap. An Australian competitor can be checked free on ASIC's registers in ten minutes and you cannot, so volunteer the equivalent information rather than waiting to be asked.

The question type that flips the preference

One practical rule falls out of the agency against tool contrast, and it is worth applying before you decide whether this finding is a problem for you. Ask whether the thing you sell has to be physically or legally present to be delivered. A service with a site visit, a local licence, a local invoice expectation or a service area is a local answer and the engines treat it as one. A product that arrives the same way wherever the buyer is sitting is a global answer, and in those readings the global answers went to foreign companies without hesitation. A supplier sitting between those two, a service delivered remotely but sold locally, has the hardest position, and the way through it is to publish the market specific page so the engines can place you on the local side of the line.

What this measurement does not establish

It is an observation across 64 readings on two engines on one date, not a controlled experiment. We did not publish matched pages from a local and a foreign domain and compare them, which is the test that would settle the mechanism, and nobody we can check has published that test either. Perplexity, Claude and Google AI Overviews were not in this run, so this says nothing about those three. And there is no per industry rate here, because the questions spanned sectors rather than sampling one.

Two hosts in the raw set are our own data provider's fetch infrastructure rather than publishers, ccn.web.tr in 8 readings and a long reverse proxy mirror of Google's developer documentation, and they are resolved back to the pages they serve before any count is taken. One other, staging.clutch.co in 2 readings, is a directory's staging environment and not something anybody intended to publish.

Questions about local preference in AI answers for Australia

Does this mean a foreign agency cannot win Australian work

It means a foreign agency with no Australian specific published work is not in the machine answer. Our page on how a foreign company wins customers without a local office sets out the six substitutes, and publishing the market specific page is the one that applies here.

Would a com.au domain fix it

The page subject is doing more work than the suffix, and the registry is the authority on what the suffix requires. auda.org.au appeared in 4 of our readings, which tells you where the engines go on that question.

Does the same preference apply to software

No, and that is the clearest part of the result. Every tool named in these readings was a foreign company, in the same answers that named only Australian agencies.

How often should a foreign supplier retake this reading

Monthly, because what is being measured is which pages the engines can reach and that changes as you publish. Our own scorecard was rewritten during the evening we read it.

Is being named the same as being cited

No, and they fail for different reasons. A page can be cited as a source while the answer names somebody else, which is why we report the two separately and never add them.

The institutional weight behind the preference

The reason an Australian answer looks Australian is visible in the sources rather than in the engines. Across our 64 blind readings on 22 October 2026, 66 distinct Australian government hosts appeared 139 times in 25 of them, which is the heaviest government signal in any of the seven country readings we took. business.gov.au in 15, the consumer regulator in 13, the communications authority in 8, the privacy regulator in 6, the statute register in 5, plus state portals from seven jurisdictions and two local councils. When the material framing a question is that institutionally Australian, the suppliers assembled alongside it are too, and the 435 Australian domains among 1,010 hosts follow from the same thing. A foreign supplier is not being excluded by a rule; it is absent from the material. The fix is a page that belongs in that set, and the measure of whether it worked is a reading rather than an argument. AI Knows Us runs a fixed question list per market across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews with the market written into each question, and reports named and cited separately. We are a foreign tool here by that measure and we appeared in zero of the 64, which we publish rather than explain.