Ask AI the same question a month apart and a different list of vendors comes back
We put the same care-software buying questions to the same four AI systems in August and again in September 2026, and a noticeably different list of vendors came back — more different than asking the same question twice on the same day. Asked a month apart, two answers to the same question shared 35% of the vendors named across them; asked twice on the same day, they shared about half. Of the three markets, aged care moved most.
A buyer who asks an AI assistant which software to look at gets a list of vendors. The question this page answers is how much of that list is still there when the same buyer asks the same question a month later — and whether the change is bigger than what you get by simply asking twice on the same day.
It is.
What we compared
The August 2026 edition asked its questions between 13 and 14 August; the September edition asked its questions on 13 September. 384 questions were asked in both. Those are the only questions compared here. Every question that itself named a vendor was left out of both sides, because a vendor named in the question is named back almost every time and would count as stability that the answer did not earn.
Each answer was paired with the answer to the same question, from the same AI system, in the same position — every question is asked three times, and the first ask in August is compared with the first ask in September, the second with the second, the third with the third. Both editions were read against the same register of tracked vendors, the one current when this page was published, so a vendor added to the register between the two editions cannot show up as change.
How the overlap is measured. For a pair of answers, take the vendors named in either answer, and count what share of them were named in both. Two answers naming exactly the same vendors score 100%; two answers with no vendor in common score 0%.
A month apart against the same day
Finding
Asked a month apart, two answers to the same question shared 35% of the vendors named across them. Asked twice on the same day, they shared about half.
| Measure | A month apart | Same day, August | Same day, September |
|---|---|---|---|
| Vendors shared, as a share of all named across the pair | 35% | 51% | 50% |
| The same vendor named first | 44% | 58–59% | 55–59% |
| Answer pairs compared | 3,247 | about 1,000 per repeat | about 1,200 per repeat |
The same-day columns are the ordinary variation: the first, second and third asks of a question on one day, compared with each other. The month-apart column is lower than both, on both measures. Only 1.8% of month-apart pairs named exactly the same set of vendors.
A second reading pools the three asks. Instead of one answer against one answer, take every vendor an AI system named for a question across all three asks in August, and the same in September — 1,231 question-and-system pairs. Pooled this way, 61% of the vendors named for a question in August were named again for it in September, and 42% of the vendors named for a question in September had not been named for it in August. On the same pooled lists, the overlap measured as above is 43%; split the questions into two halves and it comes out at 42.7% in one and 42.5% in the other, so it does not rest on a handful of questions. Pooling raises the overlap here, because each vendor has three chances to be named in a month rather than one; the 61% runs higher again because it counts only against August’s names rather than both months’. The comparison with asking twice on the same day is made on single answers: the 35% against about half.
Aged care moved most
| Market | A month apart | Same day, August | Same day, September |
|---|---|---|---|
| Aged care | 29% (1,106) | 45% (680) | 46% (811) |
| NDIS | 32% (1,207) | 50% (746) | 47% (881) |
| Allied health | 46% (934) | 59% (560) | 59% (712) |
Every market moved more across the month than within a day. Aged care moved most: the lowest month-apart overlap of the three. Allied health moved least.
By AI system, the month-apart overlap on single answers: ChatGPT 28%, Google AI Mode 33%, Google AI Overviews 39%, Perplexity 43%. None of the four held its list steady.
What this does not show
It does not say the market changed. This measures what a buyer would have seen, not why. An AI provider can change the model or the index behind an answer without announcing it. ChatGPT and Perplexity were asked through the same configured model in both editions, but the scans did not record which model the provider actually served until 16 September, after both editions had run. A change on the provider’s side is possible and is not ruled out; nothing here claims that vendors themselves moved.
Google AI Overviews rendered far fewer answers in August. On the 384 questions asked in both editions it produced 427 answers in August against 851 in September, so its figure rests on fewer questions than the other three systems’.
This comparison does not repeat in the same form. The October 2026 edition is scheduled to ask ChatGPT and Perplexity through different models, and discloses that change itself. An October-on-September version of these figures is therefore not comparable in the same way, and the figures on this page describe August to September 2026 only. If you are reading this after the October edition, the 35%, the pooled 61% and 42%, and the market table describe these two months and not the present.
One comparison, two editions, one market. Every figure here describes 384 questions about Australian NDIS, aged-care and allied-health software, asked in August and September 2026. The figures on this page were computed from the stored answers on 24 September 2026. This is a reading, not a law.