More than half of what AI reads about care software, the vendors wrote
Restated 10 September 2026. This page carried the date 13 September in four places until 11 September 2026; the restatement itself was published on the 10th and the figures are unchanged by the correction. Noted here rather than altered quietly. This page said 48.8% and called that figure “a floor, not a ceiling”, because a host only counts as a vendor’s own site if the registry already records that domain. In September the registry grew from 79 tracked vendors to 108, and 1,499 citations moved out of the unclassified long tail and into the vendor column. Nothing about the August answers changed. The same 30,178 citations were re-sorted against a registry that now recognises more of them, exactly as the original caveat said would happen. Every figure below is the new count; the figures this page carried on 22 August are named beside them so a reader can check the size of the move. The August edition’s scores are restated separately and for the same reason.
Three hosts were proposed and withdrawn. The September review put forward 32 additions; three were vendors this index had already judged and turned down before the August edition published — one on the scope test, two for having no ABN or ACN under any spelling. They were caught before publication and are recorded as exclusions rather than left to be re-proposed every month. The figures below are the 29 that stand.
One paragraph corrected 11 September 2026. The paragraph under the engine table said the two Google surfaces “cite comparatively few sources overall” — 488 distinct hosts for AI Overviews against ChatGPT’s 404 “from a third as many citations”. The table does not support that: Google AI Mode has the most citations of the four, and the host counts it compared are not in the table. It also presented leaning on vendor sites as a Google trait, when Perplexity does it too. The paragraph now says only what the table shows. No figure in the table changed.
We classified every source cited across the August 2026 edition — 30,178 citations — by who owns the page. More than half came from the tracked vendors’ own websites, from only 130 hosts. But the four engines do not read the same web at all.
An AI answer is assembled from pages the engine retrieved. Which pages, and who wrote them, is usually invisible: you see the answer, not the reading list. The index records the reading list, because every citation returned alongside every answer is stored.
So we counted them.
What we did
Every citation from the August 2026 edition — 400 buying questions, four AI systems, three runs each, 13–14 August 2026. Each cited host was resolved against the vendor registry and sorted into one of six classes. A host counts as a vendor’s own site only when the registry already records that domain as belonging to a tracked vendor; nothing was classified by hand and nothing by resemblance.
| Who wrote the page | Citations | Share | Hosts | 22 August |
|---|---|---|---|---|
| A tracked vendor’s own site | 16,224 | 53.8% | 130 | 14,725 · 48.8% |
| Unclassified — the long tail | 10,224 | 33.9% | 1,507 | 11,723 · 38.8% |
| Official (government, regulator) | 1,428 | 4.7% | 88 | unchanged |
| Directory and comparison sites | 1,379 | 4.6% | 11 | unchanged |
| Community (forums, social, video) | 731 | 2.4% | 7 | unchanged |
| Reference (Wikipedia) | 192 | 0.6% | 1 | unchanged |
One hundred and two of the 108 tracked vendors were cited at least once. Within the 130 vendor hosts the distribution is steep: ten hosts carry 43.6% of all vendor-site citations. The long tail is genuinely long and genuinely thin — 1,507 hosts, a median of two citations each, and 753 of them cited exactly once.
The move was concentrated, not spread. Only four classes could change at all, and only two did: every citation that moved came out of the long tail and went into the vendor column. Government, directory, community and reference counts are identical to the day this page was published, because none of those hosts depend on the vendor registry to be recognised.
The four engines do not read the same web
Finding
ChatGPT takes 22.0% of its citations from vendor sites. Google AI Overviews takes 61.1%. That is a 2.8× spread on the same 400 questions, asked in the same fortnight.
| Engine | Citations | Vendor site | Long tail | Official | Directory | Community |
|---|---|---|---|---|---|---|
| ChatGPT | 3,143 | 22.0% | 47.7% | 13.5% | 1.8% | 9.0% |
| Perplexity | 9,458 | 57.5% | 30.6% | 6.2% | 5.5% | 0.3% |
| Google AI Mode | 13,809 | 56.4% | 34.8% | 1.7% | 4.8% | 2.3% |
| Google AI Overviews | 3,768 | 61.1% | 27.4% | 5.0% | 3.6% | 2.8% |
Three of the four engines lean on the pages a vendor controls: 57.5% of Perplexity’s citations, 56.4% of Google AI Mode’s and 61.1% of Google AI Overviews’ come from vendor sites, against ChatGPT’s 22.0%.
ChatGPT is the outlier, and it is the outlier in every column. It takes the least from vendor sites, the most from the long tail, the most from government and regulator pages by a factor of two, and thirty-three times more from community sources than Perplexity does — 9.0% against 0.3%. It is the only engine of the four that cites Wikipedia at all.
The gap between the engines narrowed when we looked harder. On 22 August this was a 3.4× spread; on a registry that recognises 29 more vendors it is 2.8×. Every engine’s vendor share rose, ChatGPT’s by the most in proportional terms — 17.3% to 22.0%. The ordering did not change and neither did the conclusion; the size of the effect did, and it got smaller. A reading that moves when the instrument improves is worth saying out loud.
The practical reading: the kind of page that moves one engine is not the kind that moves another. Work on a vendor’s own site is aimed almost entirely at the two Google surfaces and Perplexity. It is close to invisible to ChatGPT, which is reading forums, regulators and a very long tail of pages nobody owns.
What this does not show
Being cited is not being named. These are the sources an engine returned alongside its answer; whether a vendor was named in the answer is a different measurement and it is the one the index scores. A page can be read without its subject being recommended.
More than half of the citation weight is not addressable by anyone. A vendor’s own site is the largest single class, and no vendor can add themselves to a competitor’s comparison post. More than half of what the engines read about this market is, from any one vendor’s point of view, someone else’s page and permanently so.
The long tail is unclassified, not unimportant. At 33.9% it is the second largest class and we have not characterised it beyond counting it. Some of it is trade press, some is agency marketing, some is a single blog post cited once and never again. Classifying it is the obvious next piece of work, and September made a start: 100 of its hosts were read and judged against the published inclusion rule, which is where the 29 new vendors came from. 1,507 hosts remain, and 753 of them are cited exactly once.
One edition, one fortnight, one market. Every figure here describes 400 questions about Australian NDIS, aged-care and allied-health software, asked between 13 and 14 August 2026. Engines change. This is a reading, not a law.
The classifier is a rule, not a judgement. A host is a vendor’s own site only if the registry records that domain. Vendors we do not track and secondary domains we have not recorded fall into the long tail, which means 53.8% is still a floor, not a ceiling — and the 13 September restatement is the proof. The first version of this page put the floor at 48.8%; reading 100 more hosts moved it five points. It will move again.