Alex Cannon /
Your position in AI answers changed last month. Here's what moved it
A month ago I asked ChatGPT and Gemini 24 questions a treasury team might ask, 40 times each. I wrote down every vendor they offered.
On 26 October I asked the same 24 questions again. Same wording, same 40 runs, same place. Another 1,920 answers.
104 of the 316 vendors named in September were gone. Nobody relaunched a website. Nobody got acquired. The assistants just stopped reaching for them.
If you sell treasury software, your position in these answers moved last month whether or not you did anything. This is what moved it.
The one thing the risers had in common
I can match 35 vendors to their own website and track both numbers: how often the assistant cited that website, plus how often it offered the vendor as an option.
| ChatGPT cited their own site | Vendors | Average change in how often they were named |
|---|---|---|
| Up by more than 30% | 19 | +4.1 points |
| Roughly flat | 6 | −2.7 points |
| Down by more than 30% | 10 | −2.2 points |
The pattern runs through the individual cases too.
| Vendor | Citations of their own site | Named in ChatGPT answers |
|---|---|---|
| Agicap | 56 → 559 | 6.9% → 26.7% |
| AccessPay | 40 → 210 | 2.6% → 11.4% |
| Atlar | 157 → 440 | 10.8% → 19.6% |
| Bottomline | 31 → 217 | 2.6% → 8.2% |
| Trovata | 219 → 528 | 21.1% → 30.0% |
| Cobase | 351 → 860 | 10.7% → 19.5% |
| 360T | 27 → 6 | 5.2% → 0.7% |
| Treasury4 | 71 → 12 | 5.6% → 0.5% |
| Serrala | 177 → 73 | 15.8% → 10.2% |
Thirty-five vendors is a small sample, so hold this loosely. And the direction of cause is genuinely unclear: the assistant might read your pages because it already decided to name you.
But notice which of the two numbers you can actually do something about. You cannot make a model name you. You can publish pages worth retrieving. You can check every month whether they are getting read.
Ripple Treasury is the counterexample, worth sitting with. Its citations went 268 → 533 while its naming fell 36.2% → 29.2%. Getting read is not the same as getting recommended.
The 104 that vanished
Almost all of them sat under 5% in September, which means one or two answers out of forty. That is the cheapest part of the list for a model to drop.
A name on this list means the assistants stopped offering it. It says nothing about the product.
Look for your own name:
Stratgik, Kordis, ChatFin, C8 FX Hedge, Piteco, LimFi Labs, Kognitos, HEDj, Tresora, Integral, Redwood, Graphite Connect, DebtBook, Visa, Qotid, Moneycorp, Instimatch, Centrol, CNS Treasury, ZagTrader, TreasuryFlow, Solvix, Kissflow, Iberpay, Bond Treasury, Workiva, TabaPay, Syft Analytics, Statrys, Statement, Sphere, Numeric, Moneyhub, Masttro, Flowie, Fennech, Ducasse Capital, Cynopsis, Coperitas, CellPoint Digital, Cargas, Brite, Axletree, Zodiac, Vee Fintech, Turtl, TruHedge, Treasury Nexus, Toca, TimeFX, TieSet, Taxdoo, Tackl, Sybrin, Sigma Conso, SaaScada, ProcureDesk, Paymentgate, PayU, OpusCapita, OpenCTRM, Okoora, Numo, Nortridge, Nodapay, Neonomics, Lucanet, LoanPro, LiquidityFX, KyckGlobal, HyperSwitch, Hubifi, HedgeMetrics, HedgeIn, FundCount, Flinks, Firco, Finmechanics, Finexer, Finantrix, Financial Navigator, Fabrick, FXLoader, FPtag, Elyx, EXL, DucTreasury, DataVisor, Cunexus, Crescit Pay, CorpHedge, CentraVista, Cashflo, CashLab, CashCtrl, Carlson Cash, Capdesk, Calastone, ByAllAccounts, Balancio, Avaloq, Amacor, Alemba, Addepar.
One name went the other way. Technosis was in zero of September’s 1,920 answers. It is now in 28% of ChatGPT’s in-house bank answers, 34 out of 120. Technosis AG is a Hamburg firm that has sold treasury software since 1992, so this was not a launch. It is a thirty-year-old vendor that AI had simply never mentioned, until it did.
That cuts both ways. Being absent is not permanent. Being present is not either.
The list got shorter nearly everywhere
| September | October | |
|---|---|---|
| Distinct software vendors offered | 316 | 213 |
| ChatGPT’s list | 154 | 109 |
| Gemini’s list | 277 | 200 |
| Offered by both | 115 | 96 |
| Vendors per ChatGPT answer | 5.0 | 5.6 |
| Vendors per Gemini answer | 6.7 | 7.1 |
Read those last two rows carefully. The assistants did not get stingier about how many vendors they name. They name slightly more. They draw them from a pool a third smaller.
By area, counting every vendor offered at least once:
| Area | ChatGPT | Gemini |
|---|---|---|
| Cash forecasting | 51 → 32 | 72 → 63 |
| Cash visibility | 46 → 17 | 88 → 79 |
| Bank connectivity | 32 → 23 | 64 → 53 |
| Payments | 47 → 35 | 82 → 74 |
| FX risk | 38 → 32 | 78 → 66 |
| Liquidity and investments | 31 → 20 | 40 → 43 |
| In-house bank and intercompany | 31 → 26 | 85 → 69 |
| Treasury management system | 26 → 26 | 61 → 43 |
Cash visibility on ChatGPT lost nearly two thirds of its names. If you sell there, you are now competing with 17 vendors for ChatGPT’s attention instead of 46. Easier to stand out, much harder to get in.
Kyriba is still the answer. 79% of ChatGPT answers both months. On Gemini it went up, 70% to 77%. A month of churn underneath it changed nothing at the top.
Who moved, beyond the noise
With 120 answers per area per assistant, anything under about 9 points is not a move. These cleared it.
| Area | Assistant | Vendor | September | October |
|---|---|---|---|---|
| Cash forecasting | ChatGPT | Agicap | 25.8% | 90.0% |
| Treasury management system | ChatGPT | Agicap | 21.7% | 70.0% |
| Bank connectivity | ChatGPT | AccessPay | 9.2% | 55.8% |
| Cash visibility | ChatGPT | Agicap | 6.7% | 52.5% |
| FX risk | ChatGPT | Milltech | 2.5% | 41.7% |
| Payments | ChatGPT | Nomentia | 28.3% | 66.7% |
| Bank connectivity | ChatGPT | TIS | 40.0% | 71.7% |
| In-house bank | Gemini | Corpay | 1.7% | 35.8% |
| In-house bank | ChatGPT | Serrala | 43.3% | 5.0% |
| FX risk | ChatGPT | 360T | 40.8% | 5.8% |
| FX risk | Gemini | Ripple Treasury | 61.7% | 31.7% |
| In-house bank | ChatGPT | Ripple Treasury | 68.3% | 39.2% |
| FX risk | ChatGPT | FIS | 60.0% | 31.7% |
| Treasury management system | ChatGPT | FIS | 69.2% | 41.7% |
Agicap went from a quarter of ChatGPT’s cash forecasting answers to nine in ten. In one month.
Serrala and 360T are the ones I would study if I worked at a vendor. Both held a strong share in exactly one area. Both lost almost all of it. A position that rests on one question is thinner than it looks on a dashboard.
Gemini stopped arguing against the big platforms
September’s strangest finding was Gemini naming an enterprise platform purely to warn the reader off it. On the question about a three-person treasury team, it named Kyriba in 17 of 40 answers only to rule it out.
| Assistant | Vendor | September | October |
|---|---|---|---|
| Gemini | Kyriba | 18 | 5 |
| Gemini | FIS | 11 | 2 |
| Gemini | ION | 6 | 1 |
| Gemini | SAP | 3 | 0 |
Across the study, rule-outs fell from 43 answers to 14.
On that small-team question Gemini now offers Kyriba as a real option in 32 of 40 answers, up from 23. In three answers it skips Kyriba entirely.
If you sell against the enterprise platforms, you lost an ally. A month ago one assistant was making the “too heavy for you” argument unprompted. You have to make it yourself now, on your own pages, in words a model can quote.
Both assistants are reading twice as much
| Cited sources | September | October |
|---|---|---|
| ChatGPT | 7,256 | 13,409 |
| Gemini | 2,612 | 4,380 |
ChatGPT cites about 14 sources per answer now, against 8 in September, across 401 domains rather than 353.
So there is more room to be read and less room to be named. That combination is the whole story of this month.
One more number for anyone who screenshots an AI answer and sends it to their board: ask ChatGPT the same question twice and the two shortlists now share 51% of their names, up from 41%. Gemini is at 34%. Fewer than one Gemini pair in two hundred gives exactly the same list. A single screenshot still tells you almost nothing.
What to do about it this week
Count your own citations, not your mentions. Pull the sources panel on ten answers to questions your buyers ask. Count how many point at your domain. That number moved with naming for 19 of the 35 vendors I could track. It is the half you control.
Find the pages doing the work. For Agicap, Nomentia and Cobase the citations land on specific pages, not homepages. Look at which of your URLs get cited and write more of that, rather than more of everything.
If you dropped off the list, do not start with rankings. You are not losing a race, you are absent from it. That is a retrieval problem: thin pages, gated content, nothing a model can quote.
Check more than one area. Serrala and 360T both looked strong on one question and fell off a cliff. If your whole position sits in one area, you are one model update from Treasury4’s month.
Run the 24 questions yourself. They are published at the bottom of the September piece, word for word. Ask them five times each in both assistants, write down who gets offered and keep the file. It takes an afternoon. Next month you will have your own before-and-after instead of mine.
And if you are a treasurer rather than a vendor: ask both assistants, ask more than once, then ignore the ranking and read the sources. They only agree on 96 vendors out of 213.
Something changed inside ChatGPT
In September ChatGPT reported itself as gpt-5-6. In October it reports gpt-6 across all 960 answers. Gemini reported 3.5 Flash-Lite both months.
ChatGPT’s half of this study therefore compares two different models. Whether Agicap climbed because of anything Agicap did, or because the machine underneath changed, is not something these numbers can separate.
Gemini held its model steady and shrank much less: 277 vendors to 200, against ChatGPT’s 154 to 109. Some of the contraction is probably the model rather than the market. I would not bet on how much.
This is the uncomfortable part of measuring AI visibility at all. The ground moves under the measurement. You only find out because you wrote down the model string.
How it was measured
The same 24 questions as September, checked character by character against the file I locked on 19 September. 40 runs each on ChatGPT and Gemini, web search on, from the UK in English, on 26 October 2026. All 1,920 answers came back with text.
Both months were scored by the same code, which decides whether a vendor was offered as an option or merely mentioned in passing, ruled out or named as part of the reader’s own setup. Run over September’s answers it reproduces the figures I published in September, which is how I know the two months are measured the same way.
Two vendors joined the name list for both months, after I checked what October named that September’s list could not see: Technosis and CompuHedge. I went through every candidate name appearing in two or more October answers. Those were the only real companies in there.
The rules and the threshold for what counts as a change were written down and committed to git before any data was collected.
I will run these questions again in late November. If you want your own name tracked properly, the honest answer is to run them yourself and keep the file.
Questions, corrections or a vendor I have miscounted: alex@alexcan.xyz.