Research · August 2026 · Ployo
The Unnamed Category
We asked ChatGPT, Perplexity and Google AI Overviews 235 questions an Australian care employer would actually ask about hiring. In 78% of the answers, they recommended no one at all.
Why we ran this
Care employers have started asking AI where to start
Australian health care and social assistance is the country's largest industry by employment and carried 55,300 job vacancies in August 2025, close to double retail trade, according to the ABS. The NDIS Review reports that between 17% and 25% of disability support workers leave their job every year.
People running those rosters are increasingly typing their problem into an AI assistant before they type it into a search engine. Nobody had measured what comes back. So we did.
The short version: the answers are careful, heavily sourced from regulators, and almost entirely empty of recommendations. Where recommendations do appear, the three engines rarely agree with each other.
Findings
Five things the data shows
Most answers name nobody
Of 705 answers to Australian care hiring questions, 548 named not a single product, platform or supplier. The engines answered with advice, process and regulation instead. Only 157 answers named anyone at all, and across the whole set just 36 distinct vendors were named even once.
The three engines almost never agree
When we compared the sets of vendors each engine named for the same question, the mean overlap was 0.09 on a scale where 1.0 is perfect agreement. All three engines named at least one vendor on only 29 of 235 questions. Of the 48 questions where two or more engines recommended anyone, only 27% shared a single name in common.
Practically: ask ChatGPT and Perplexity the same hiring question and you will usually get two entirely different shortlists.
How you phrase it decides whether you get recommendations at all
We wrote half the questions as shopping ("best AI screening software for aged care providers") and half as problems ("candidates keep not showing up for support worker interviews"). Same topics, same country, same engines.
Shopping questions produced a named vendor 43% of the time. Problem questions produced one 4% of the time. An employer who describes their situation rather than naming a product category will be given advice and no options.
Regulators are the backbone of the answers
A quarter of the 6,103 citations with a resolvable destination came from Australian government domains. The NDIS Quality and Safeguards Commission was the single most cited source in the entire study, followed by the Department of Health, AHPRA and Fair Work. Reddit and LinkedIn were the largest non-government sources.
On compliance questions the pattern is starker still. Australian government domains account for 57% of those citations, led by AHPRA at 13.3% and the NDIS Commission at 7.2%, and a vendor is named in only 11% of the answers.
Almost nothing is settled
We counted a question as "owned" when a single vendor was named by all three engines. That happened on eight questions out of 235. Ployo and HireVue owned two each. Employment Hero, Humanforce, Paradox and Sapia.ai owned one each. On a looser test, a vendor named by any two of the three engines, it rises to 29 questions, still under an eighth of the set.
For a category that has existed for a decade, there is remarkably little consensus for an AI to inherit.
Who gets named
The answers are not dominated by interview software
Share of the 235 questions where each vendor was named by at least one engine. Read this as visibility inside AI answers, not as market share, revenue or customer count.
Five of the fifteen most-named names are not interview or screening products at all. Employment Hero, ShiftCare, Worknice, Classet and JobAdder are HR suites, rostering platforms and applicant tracking systems. When an Australian care employer asks an AI how to fix their hiring, the answer often points at the system they already run.
Cited or recommended
Being read is not the same as being recommended
Naming and sourcing are two different things. An engine can read a company's pages, attach them as a source, and still not put that company in the answer. Ranked by how many times a domain was cited across all 6,103 resolved citations, the top of the list is almost entirely regulators and public forums.
| Rank | Domain | Citations | Questions | What it is |
|---|---|---|---|---|
| 1 | ndiscommission.gov.au | 232 | 56 | Regulator |
| 2 | health.gov.au | 185 | 70 | Government |
| 3 | ahpra.gov.au | 174 | 25 | Regulator |
| 4 | fairwork.gov.au | 140 | 50 | Regulator |
| 5 | reddit.com | 125 | 53 | Forum |
| 6 | linkedin.com | 118 | 76 | Social |
| 7 | pmc.ncbi.nlm.nih.gov | 81 | 43 | Research |
| 8 | agedcarequality.gov.au | 73 | 36 | Regulator |
| 9 | ployo.ai | 71 | 50 | Vendor |
| 10 | oaic.gov.au | 69 | 45 | Regulator |
The first vendor of any kind appears at rank 9, and it is ours. That is a disclosure, not a boast, and the number underneath it is the more interesting one. ployo.ai was cited in 50 of the 235 questions but named in the answer in 28. The engines read the material and then wrote an answer that did not mention where it came from. Every other vendor sits further down: employmenthero.com at rank 16, sapia.ai at 22, shiftcare.com at 29, paradox.ai at 46, hirevue.com at 73. HireVue is the most-named vendor in the study and the 73rd most-cited domain, which is a reminder that the two measures are not the same thing and that a name can travel through these systems without a source attached to it.
By question type
Recommendations cluster in a narrow band
| Question type | Answers | Named a vendor |
|---|---|---|
| Which tool should I use | 198 | 45% |
| Does it integrate with what we run | 18 | 39% |
| Is it any good, who else uses it | 30 | 37% |
| What does it cost | 30 | 33% |
| How does A compare to B | 27 | 26% |
| What are my other options | 21 | 19% |
| How do I do this | 42 | 12% |
| Is this legal, what are the rules | 132 | 11% |
| Here is my problem, help | 207 | 4% |
The two largest groups sit at opposite ends. Questions that already assume a product category get a shortlist. Questions that describe a workforce problem, which is how most people start, get advice.
By engine
Three engines, three different personalities
| Engine | Citations per answer | Answers naming a vendor | Vendors named per answer |
|---|---|---|---|
| Google AI Overviews | 2.1 | 28% | 0.60 |
| ChatGPT (web search) | 4.5 | 22% | 0.52 |
| Perplexity (Sonar) | 19.6 | 17% | 0.37 |
Perplexity reads roughly nine times as many sources per answer as Google and recommends vendors least often. Google reads the fewest and recommends most. More sourcing does not produce more recommendations. If anything it produces fewer, because a well-sourced answer to a workforce question is a regulator's answer, and regulators do not name suppliers.
Disclosure
Ployo built this study and Ployo sells AI interview software in this market. Ployo appears in the data at 11.9%, second to HireVue. We did not write the questions: they were generated by language models given only a buyer persona and no knowledge of Ployo or any other vendor, then screened to remove any that named a product. The full question set, the raw answers and the citation list are available to download below so that anyone can reproduce or contradict this.
The finding least convenient to us is in the data too. Across the 372 problem-phrased answers, where most real buyers start, Ployo was named in one. Almost nobody was named in any of them.
Method
- Questions
- 235 queries about Australian care sector hiring, covering registered and enrolled nurses, assistants in nursing, personal care assistants, disability and NDIS support workers, home care workers, allied health and mental health roles, plus operational questions about volume, turnover and rostering. Written by six language model agents given an occupation cluster and a buyer persona, with no information about any vendor. Any query that named a product was discarded before the run.
- Balance
- 111 shopping-phrased, 124 problem-phrased. Nine question types. Deliberately mixed vocabulary: some queries say "AI interview", some "phone screening", some "automated screening", some name no technology at all.
- Engines
- OpenAI with the hosted web search tool, Perplexity Sonar, and Google AI Overviews retrieved through a search scraper. All three run search-grounded. Ungrounded model calls were excluded because they measure model memory rather than what a user sees.
- Location
- All queries run with Australian geo-targeting. Fieldwork 26 August 2026, with one question re-run on 28 August 2026.
- Citation counting
- 6,150 citation entries were returned in total. 6,103 resolved to a destination domain; 47 were relative Google redirect URLs whose target could not be recovered and are excluded from all domain-level figures. Domain percentages use 6,103 as the denominator.
- Counting
- A vendor counts as named when its name appears as a whole word in the answer text. Vendor names that are ordinary English words were excluded from counts to avoid false positives. Citations are the sources each engine attached to its own answer.
- What the vendor count can and cannot see
- Answers are matched against a fixed list of 58 vendor names assembled before the run. 36 of them were named at least once. A product outside that list is invisible to this count, so the 36 is a floor rather than a ceiling, and a vendor's share here is a share of the questions we asked, not of a market. The list was built to over-cover Australian care sector suppliers specifically, which is why it names rostering, care management and HR platforms alongside interview software. Names are counted as the engines returned them. We did not verify that every name corresponds to a product that exists, is still trading, or sells into this market, so a name appearing here is evidence about the answer, not about the company.
- Agreement
- Engine agreement is a Jaccard index computed across all three engines at once, not pairwise: for each question we divide the number of vendors named by every engine that named anyone by the number named by any of them, then take the mean. It is computed only over the questions where at least two engines named at least one vendor. On the same basis, a question counts as shared agreement when that all-engine intersection is not empty.
- Scale
- 705 answers in the main sweep. A further 895 answers across two earlier sweeps, including a 48-question blind pilot, informed the design and are cited where referenced.
Limits
What this study cannot tell you
- It measures what AI engines say, not what employers buy. Nothing here is market share, revenue, customer count or product quality.
- It is a single point in time. These systems change weekly and the same questions asked next month may return different names.
- Question volume is not demand. We wrote 235 questions because they cover the ground, not because Australians ask them in that proportion.
- One country, one sector. Nothing here should be read as applying to hiring generally or to markets outside Australia.
- Three engines out of many. Copilot, Claude, Gemini in its chat form and every assistant embedded in another product were not tested.
- Vendor naming is measured, sentiment is not. An answer that names a vendor to criticise it counts the same as one that recommends it.
What we think it means
An empty answer is not a neutral one
The instinct on seeing 78% is that AI is being cautious, and that is partly true. But an employer who asks how to fix a support worker shortage and receives a well-cited summary of the NDIS Code of Conduct has not been helped much. The advice is correct and the question is unanswered.
For the care sector specifically this matters more than in most industries, because the people asking are usually not procurement professionals. They are facility managers and rostering coordinators doing recruitment on top of another job. The gap between "here is the regulation" and "here is what other providers do" is exactly where they are stuck.
For anyone selling into this market, including us, the finding is uncomfortable and simple. There is no incumbent to displace. There is an absence, and it is being filled by regulators because nobody has written anything better.
Download the data
The question set, the raw answers, and every citation
Four CSVs: the questions we asked, every answer each engine gave, every domain cited, and the vendor-share rollup behind the chart above.
Published under a Creative Commons Attribution 4.0 licence. Use it, republish it, check our working, and cite Ployo as the source. CC BY 4.0.