What 988 Australian therapy websites are missing

The four things most practices are missing

Page labels, on 88.5% of practices. At least one missing or duplicated title, heading or description across the interior pages we sampled. Most often no main heading at all (195 practices) or no page description (184). These are the labels a search engine and an AI answer read first to work out what a page is about, and they are the cheapest thing on this list to fix.

No llms.txt, on 66.4%. A proposed standard for telling AI assistants what a site offers. We report it because it is widely discussed and rarely measured, and we deliberately do not mark anyone down for lacking one: there is not enough evidence that having the file changes whether a practice gets recommended, and Google states it is not required for Search or its AI features. If someone offers to sell you an llms.txt file as an AI fix, that is the tell.

No structured data about the practice, on 24.8%. No machine-readable record of the business, its clinicians or its services. A further 176 practices publish some but leave the identifying parts empty, which takes the figure to 44.5% if partial counts as missing. Some carry theme furniture instead, page headers and menus marked up as data, describing nothing about the practice.

Blocking AI crawlers, on 3.2%. robots.txt refuses GPTBot, ClaudeBot, CCBot, Google-Extended and others by name, or disallows everything. The smallest number here and the only one that directly prevents access rather than making it harder. The identical block list recurs across unrelated practices in different states, so it arrives with a plugin rather than a decision.

By state and territory

Nationally, 57.7% of the 964 practices we could grade fell below our website readiness threshold, which asks whether crawlers are allowed in, whether the site is indexable, whether it publishes structured data about the practice, and whether it has a working sitemap. Weighting jurisdictions by population rather than by how many practices each contributed gives 56.9%.

The practices Google's AI names are better built than the rest

Everything above describes how sites are made. This is the part that connects it to being found. We asked Google's AI Overviews and AI Mode a patient's question in all 102 locations, twice over: "I have been struggling with anxiety and I think I need to see a psychologist in Newcastle, New South Wales, Australia. Who would you suggest I contact?" Then we took every practice those answers named and measured it the same way we measured the random sample.

A 4.3 point gap, and it is statistically significant (z 2.23, p about 0.025). The direction holds on every measure we have: the practices Google named publish more text in their page source, 102 words more at the median, and are roughly a third as likely to hide it behind JavaScript.

How much text a practice publishes in its page source

This is the narrowest measurement here and the easiest to overstate, so here is exactly what it is. We fetched each practice's home page only, did not run any JavaScript, removed the menus, header and footer, and counted the words of text left behind. Most AI crawlers do not run JavaScript, so this approximates what they receive when they arrive.

The median practice publishes 464 words of text, with quartiles at 248 and 785. A tenth sit under 121 words and a tenth above 1,297. 82.9% carry at least 120 words, which we treat as the least a page needs to state what it treats, where it is and who works there.

The 17% we could not read, and why

Three distinct failures hide behind a single word count, and treating them as one thing would be the whole error. Only one of the three is about a practice writing too little.

Bot protection answered instead of the site, 8%. The host returned a challenge page rather than the practice's content. One answered in 168 bytes. If an ordinary browser request cannot get through, a crawler is unlikely to do better. A hosting default, not a decision the practice made.

Genuinely sparse, 6.1%. The page loads and carries real content, there is just very little of it: under 120 words once navigation is removed.

JavaScript-only, 3%. The sharpest failure in the study. One practice served 517 kilobytes of markup containing 40 words of text, every one of them a menu label: "Home", "How we can help", "Contact Us", "Use tab to navigate through the menu items". The content exists, but only once a browser runs the page's JavaScript. That practice has written everything a patient needs and a non-rendering crawler receives a navigation bar.

Does our own grade mean anything?

Our free scan gives every site a readiness grade from how it is built. That grade asserts a site is easier for a machine to read, and nothing in this study had tested the claim, so we cross-tabulated the two. Practices below our readiness threshold were 13.0% unreadable. Those above it were 6.4%. A 6.6 point difference, in the direction the grade claims. A modest relationship, not a strong one, and we would rather publish it than assert the grade works: the threshold carries real signal and is nowhere near sufficient on its own.

Registered and unregistered practitioners

Australia regulates these professions differently and we wanted to know whether it shows. Psychologists register with AHPRA, hold a protected title, and cannot use testimonials in advertising they control under section 133 of the National Law. Counsellors and psychotherapists are not AHPRA-registered, are not bound by that ban, and cannot offer a Medicare rebate, so they compete for private-pay clients on different terms.

We tagged each practice by the search term that surfaced it: 294 graded through psychologist terms, 325 through counsellor and psychotherapist terms, and 345 matching both, which were set aside. 60.2% of the psychologist-term practices fell below the readiness threshold, against 55.4% of the others. That 4.8 point gap is not significant (p is about 0.23). Before the fieldwork we set the smallest gap this sample could resolve at 6.5 points and recorded that anything under it would prove nothing either way. This is under it, so we could not answer the question, which is not the same as finding the two groups alike. Settling it needs a check against the AHPRA register, not a larger sample.

Methodology

988 independent practice websites, drawn from 408 Google Places queries across 102 locations: metropolitan suburbs, because Places returns a different business set for Bondi Junction than for Sydney, plus regional centres so the sample is not purely capital-city. Every query returned. We excluded everything that is not an independent practice website: booking platforms and practice software, professional-body directories, telehealth marketplaces, insurers and hospital groups, national charities and crisis services, general business directories, NDIS aggregators, social profiles, and government and education domains. We also excluded every practice with any relationship to Moonraker, so none of our own clients or leads appears here.

Each practice was scanned on 25 and 26 September 2026 with the engine behind our free scan, one version throughout, then fetched again independently for the text measurement with an ordinary browser user agent. We used a browser agent rather than GPTBot's because the question there is what the HTML contains; whether a host serves it to GPTBot at all is the robots.txt finding above.

What this study cannot tell you

We read one page per practice. The text measurement is the home page only. A practice with a thin home page and excellent service pages counts as thin there, and the page-label findings come from a different check with its own sample of pages. Nothing here describes a whole website.

We measured who gets named, not what naming rewards. We asked two Google AI surfaces and did not ask ChatGPT, Claude or Perplexity, so "AI" here means Google's answers and nothing broader. The comparison shows better-built practices are over-represented among those named; it cannot show that improving a site would get a practice named.

Re-running our sampler will not reproduce our sample. The shuffle is seeded, but the search results underneath it are not stable. Two identical runs an hour apart shared only 67% of their practices. Anyone repeating the method gets a comparable sample, not an identical one.

The profession split is a search term, not a register check. Comparing two runs, 10.9% of practices landed in a different group the second time. Misclassification of that kind pushes any real difference toward zero, so a gap we report is understated rather than overstated.

The sample leans toward practices already visible in Google Places in larger centres, so it is not representative of every Australian practice, particularly online-only ones. And we do not publish per-practice results: naming which practices scored what would be a list of businesses ranked by a weakness they did not ask to have measured.

The dataset

Jurisdiction-level aggregates are public under CC BY 4.0: download the CSV. Cite as "Moonraker, Australian Therapy Websites, September 2026", linking this page. Journalists: for additional cuts or comment, contact scott@moonraker.ai.

These figures are not comparable with our 2026 US study, the AI Invisibility Index. That study ran on an earlier version of the grading rubric and carries its own methodology correction. Setting the two headline percentages side by side would compare different definitions, not two countries.

If you run a practice

Check your own site with the free AI readiness scan, or see whether the rest of the web agrees about your details with the listings check. Both are free and neither needs a conversation with us.