Which AI Engines Name Australian Clinics

Across 12 Australian patient questions put to four AI engines on 12 September 2026, clinic and practice websites took 60% to 72% of all citations, and directories such as HealthEngine took only 6% to 21%. The catch is overlap. Of the 167 clinic websites cited, 74.3% were cited by exactly one engine and only six were cited by all four. Google declined to show an AI Overview at all for 3 of the 12 questions, including the most commercial one in the set.

Which AI engines actually name Australian clinics?

All four engines name clinics, and they do it far more readily than most practice owners expect. What differs is how often, and how widely they spread the naming around.

On 12 September 2026 we put the same 12 questions to four AI answer surfaces. The questions are the kind a patient actually types, in four groups of three: pick me a provider in a named suburb, here is my symptom and my city, here is a specific service I need, and three general questions with no location at all as a control.

Every answer was stored, and every citation in it was reduced to a domain and classified. Here is how often each engine cited a provider's own website in its answer.

Perplexity cited a clinic in every answer. Google AI Overviews managed 8 of 12, partly because it refused to produce an overview at all three times, which is its own finding further down.

Across all four engines, 167 distinct Australian clinic and practice websites were cited. Being named by an AI engine is not a rare prize. It is the normal state of affairs.

The breadth is where the engines part company. Perplexity cited 125 distinct provider sites across its 12 answers. ChatGPT cited 31.

That gap matters commercially. On Perplexity a clinic is one of a crowd of 125. On ChatGPT it is one of 31. The same visibility work does not buy the same thing on each.

One engine is missing from this table and we are not going to pretend otherwise. Gemini was asked all 12 questions and returned an API credit error every time, so it is recorded as not assessed. It is not a zero. We could not ask, which is a different thing from being invisible.

Do AI engines send people to directories or to clinic websites?

Clinic websites, by a wide margin. This is the result that contradicts the advice most clinics are given.

The standard line is that AI search runs on directories, so a practice should pour its effort into HealthEngine, HotDoc, Cleanbill and the health insurer finders. We classified every single cited domain to test it.

Google AI Overviews cited a directory in 6.2% of cases. It was the least directory-dependent of the four, not the most.

Perplexity leaned on directories hardest at 21.3%, and it still cited clinic websites three times more often than directories.

The practical reading is narrow and worth stating carefully. Directory listings may be worth keeping for booking, for reviews and for patients who search that way. They were not the thing these four engines mostly quoted. The page the engines quoted was the clinic's own.

If one engine cites your clinic, do the others?

Usually not. This is the finding with the most uncomfortable implications.

Of the 167 clinic websites cited across the four engines, here is how many engines cited each one.

124 of 167, or 74.3%, were cited by exactly one engine. Six sites were cited by all four.

So "we show up in AI search" is not a single status a clinic has. On this evidence it is four separate races, and winning one says almost nothing about the other three.

There is a genuine counter-current, and leaving it out would misrepresent the data. Look at a single question rather than at the whole pool, and the engines agree far more than the 74.3% suggests. In 11 of the 11 questions where at least two engines cited any clinic at all, at least two of them cited the same clinic.

Both things are true at once, and the combination is the actual finding. Every question has a small shared core of clinics that most engines land on, surrounded by a large tail that is unique to each engine. A clinic in the core is visible everywhere. A clinic in the tail is visible in one place and has no way of knowing it.

Why did Google refuse to answer the most commercial question?

Because it is the most commercial question, and Google is most careful exactly there.

Three of the 12 questions produced no AI Overview at all. Not a thin one. None.

The pattern is not proven by three cases and we are not going to claim it is. But the three are suggestive, and they fit what Google says about its own health handling.

"Best physiotherapist in a named suburb" asks Google to rank named health practitioners against each other. "Where can I get a skin cancer check" asks it to direct someone toward investigating a possible cancer. Both are squarely in the territory Google treats as high-stakes, and on both it produced nothing and left the ordinary results to do the work.

For a clinic, that is useful rather than disappointing. On the queries closest to a booking, the classic blue link was still the entire answer. Ranking has not stopped mattering.

What happens when an engine gets the country wrong?

It sends the patient to the wrong hemisphere, and it looks completely confident doing it.

Seven citations in this run pointed at organisations outside Australia. All seven came from a single engine, Perplexity, and they clustered in two answers.

The Newcastle one is the clearest. There are two famous Newcastles and the engine picked the English one, despite the question saying New South Wales.

This is the quiet argument for putting your suburb, your state and your country in plain text on your own pages. Not in a schema block a human never sees. In the sentences.

An engine that cannot tell which Newcastle you are in will not guess in your favour.

How was this measured, and what did the machine get wrong?

Openly, and the machine got about a fifth of it wrong before a human fixed it.

The questions. Twelve, written before any data was collected, in four groups of three: three asking for a provider in a named suburb, three giving a symptom and a city, three naming a specific service and a city, and three general questions with no location as a control.

The engines. ChatGPT through the OpenAI Responses API with web search and an Australian user location, Claude Sonnet 4.5 with a web plugin, Perplexity Sonar, and live Google AI Overviews captured from the real Australian result page. Every engine got the same 12 prompts and the same neutral system instruction. Every raw answer was stored.

The classification. Every cited URL was reduced to a domain and put in one of four buckets: a provider's own website, a directory or comparison site or insurer, a peak body or government or research source, or unverified.

The mistake, and this is the part worth reading. The first pass was automatic. Any domain not on a known directory or government list was counted as a clinic website. That over-counted badly. A human then opened all 212 flagged domains, read the site's own homepage title and description, and moved the ones that were not a clinic.

45 of 212 were wrong, a 21.2% error rate on the machine pass. YouTube was in there. So was PubMed. So were three dictionaries, a university, a consumer advocacy group, and 23 directories and comparison sites the keyword list had simply never heard of.

Every number in this post is the hand-corrected version. We are publishing the error rate because any tool selling you an automated "AI visibility score" is running the pass we ran first, and is very unlikely to be telling you how often it is wrong.

Nine domains would not load at all and are counted as unverified rather than guessed at.

We have deliberately not published the names of the Australian clinics that were cited. Reproducing a list of named practices reads as a recommendation, and a measurement of what a language model said this week is not a recommendation.

What does this change for a clinic?

Three things, and the first one is the cheapest.

Check it yourself, per engine, and write down what you find. Ask each engine the three or four questions your patients actually ask, in the wording they use, and record which clinics come back and what gets cited. Do it on all four, because 74.3% of cited clinics appeared on only one. A single check on a single engine tells you close to nothing.

Put the effort into your own pages before the directories. Between 60% and 72% of what these engines cited was a provider's own website. If the choice is between improving a service page and topping up a directory profile, this data points at the service page.

Say where you are, in words. The country errors in this run were not subtle. Suburb, city, state and country belong in the visible sentences on the page, not only in the schema.

And a caution on the method. This is 12 questions on one day, and engines change weekly. The value is in the procedure, not in these exact percentages. Re-run it next quarter and the numbers will move. The point is that you can re-run it at all, which is more than can be said for most claims made about AI search.

Method and sources. 12 prompts × 4 engines, run 12 September 2026. ChatGPT via the OpenAI Responses API (gpt-5-mini, web_search tool, user location AU); Claude Sonnet 4.5 and Perplexity Sonar via OpenRouter; Google AI Overviews captured live from the Australian result page. Gemini returned an API credit error on all 12 prompts and is recorded as not assessed. 427 citations classified in total; 212 domains hand-reviewed against their own homepage titles, of which 45 were reclassified. Raw answers, the domain evidence and the classification were kept so the run can be repeated.

Related reading:

• How to rank in AI search
• When Google answers your patient's question for you
• What patients read on a clinic's About page before they book
• SEO and GEO fix for clinic websites
• View all services →