Alex Cannon /
What ChatGPT and Gemini tell treasurers about 58 treasury vendors
A treasurer hears a vendor’s name at EuroFinance or sees it on a shortlist and instead of finding the website, they might just ask ChatGPT or Gemini about it.
So I did the same for all 58 treasury software vendors on Treasury Brief. Two questions, five times each, on both assistants, with web search on. That’s 1,160 answers.
The short version: the two assistants describe the same vendors from very different pages. That changes what they tell the buyer.
| ChatGPT | Gemini | |
|---|---|---|
| Sources that were the vendor’s own pages | 90% (3,371 of 3,768) | 33% (531 of 1,617) |
| Answers built mostly on other people’s pages | 11% (61 of 570) | 64% (336 of 523) |
| Answers about the wrong company | 1% (8 of 580) | 5% (27 of 580) |
| Answers that got the buyer wrong | 11% (24 of 210) | 20% (43 of 210) |
| Answers that told a real buyer “it’s not for you” | 0% (0 of 210) | 20% (42 of 210) |
ChatGPT describes a vendor from the vendor’s own site. Gemini mostly describes it from everyone else’s.
Where the answers came from
With web search on, each assistant picks its own pages. I sorted every page they cited by who published it.
| Source | ChatGPT | Gemini |
|---|---|---|
| The vendor’s own pages | 90% | 33% |
| Review and list sites | 3% | 23% |
| Competitors’ pages | 1% | 9% |
| Consultants and partners | 1% | 7% |
| News and press | 1% | 5% |
| Other (LinkedIn, company databases, marketplaces) | 5% | 23% |
Coupa is a good example. I asked Gemini what Coupa is five times and it didn’t cite a single Coupa page. Every answer leaned on a procurement firm’s explainer and on “What is Coupa used for?”, a page published by Zip, a competitor.
I expected smaller vendors to get described from other people’s pages more than the big ones. They didn’t. Who a vendor sells to barely mattered: 68% of sources were the vendor’s own for enterprise-only vendors, 73% for everyone else.
What mattered was which assistant the treasurer asked.
Gemini tells buyers a vendor isn’t for them
The second question was: “What kind of company is it built for, and who would it not be a good fit for?”
ChatGPT hedges here. It’ll say a tool isn’t for “a small business with simple finances”, but it never rules out small businesses as a whole.
Gemini is blunter. In 42 of its answers, it told a group of buyers the vendor actually sells to that the vendor wasn’t for them. ChatGPT didn’t do this once. The group Gemini turned away most often was mid-market companies.
HighRadius is the clearest case. It sells to mid-market and enterprise finance teams. All 5 Gemini answers told mid-market companies it wasn’t for them.
One called it “massive overkill”. None of the 5 cited a HighRadius page. Three cited competitors’ “HighRadius review” pages instead.
It happened in all 5 Gemini answers for AccessPay, Datalog, ION Treasury, Nomentia and Workday too.
Both know what a vendor sells. Neither is sure who it’s for
The first question was: “Someone on our finance team mentioned [vendor]. What is it and what does it actually do?”
Not one answer got the product wrong or left it vague.
The buyer is where things slip. For the 42 vendors whose site says who they sell to, about one answer in six named the wrong buyer.
And the misses nearly all went the same way. The answer made the buyer wider than the vendor claims. In 63 of 67 wrong-buyer answers, the vendor’s real buyer was named, plus one more. Usually the extra one was mid-market.
Enterprise-only vendors got it worst. For the 9 whose site sells only to enterprises, 51 of 90 answers also pitched them at mid-market companies.
If a vendor’s site doesn’t say who it’s for, the assistant decides anyway. For the 16 vendors that don’t say, the answer named a company size in 135 of 160 answers. The most common guess was “mid-market and enterprise”.
That lines up with what the sites leave out. It doesn’t prove the missing line causes the guess.
The name is a bigger risk than the headline
I expected vendors with plain homepage headlines to be described better than vendors with slogans. Wrong again: 92% right for plain headlines, 97% for slogans.
What really decided the “what is it” answer was the name. 8 vendors got mixed up with another company or thing at least once:
- Predicta: 9 of 20 answers were about Predicta Global, Predicta Analytics or PredictAP
- Treasury Systems: 7 answers explained treasury management systems in general
- Float: 5 were about Float.com or Float Financial
- Bond Treasury: 4 were about Treasury bonds or treasury departments
- QLM: 4 were about something else called QLM, mostly “Quantitative Language Models”
- Statement: 4 were about a different company called Statement
Gemini made this mistake more than three times as often as ChatGPT.
Old names stick
GTreasury became Ripple Treasury after Ripple bought it in late 2025.
Ask about “Ripple Treasury” and the answers usually explain the history. Ask about “GTreasury” and most describe GTreasury as if nothing happened. Only 2 of 10 ChatGPT answers told the buyer it’s now Ripple Treasury. Gemini managed 5 of 10.
New owners mostly go unmentioned too. Statement’s answers named its owner, Tipalti, in 4 of 9 on ChatGPT and 0 of 7 on Gemini.
So a treasurer who heard the old name from a peer walks away with the old story.
Does any of this matter?
This isn’t the end of the world. I’m sure most people look at other sources, or ask a follow-up, before trusting the first answer they get.
What it does show me is that every assistant handles the same question differently. If you’re a vendor, you can’t have one “AI visibility” plan. You need to know how each assistant builds its answer and work on that.
If you’re a treasurer
Trust “what it does” more than “who it’s for”. The product was right every time. The buyer was the weak spot.
On Gemini, open the sources behind “not a good fit”. They’re often review sites and competitors’ comparison pages, not the vendor. If the answer matters, ask ChatGPT too. For 9 vendors the two gave different answers most of the time.
If you’re the vendor
Run the same two questions on your own company:
Someone on our finance team mentioned [your company]. What is it and what does it actually do?
We're looking at [your company] for our finance team. What kind of company is it built for, and who would it not be a good fit for?
Ask each five times on ChatGPT and five times on Gemini, with web search on. Open the sources every time.
- ChatGPT gets the buyer wrong: it’s reading your site. Say who you’re for on your homepage and product pages, in plain words, with a company size.
- Gemini gets it wrong or turns away a buyer you want: your site is only a third of what it cites. Look at the review profiles and comparison pages it quoted. That’s where the work is.
- If it describes a different company, put your category next to your name wherever it appears, on your site and in directory listings.
- If it uses your old name, update your own old pages first. Then ask the review sites and directories that still carry it.
How it was measured
The 58 vendors are every vendor listed on Treasury Brief on 19 September 2026. I maintain that directory. It’s the answer key here because every fact on it links back to the vendor’s own page.
If you spot something wrong or want your company added, email me at alex@alexcan.xyz.
Both questions were asked word for word, as above. Web search was on. Each question ran five times per vendor on each assistant, from the UK, on the 19th of September.
Each “what is it” answer was scored on its opening description. Each “who is it for” answer was scored on the company size it named: small, mid-market or enterprise. Each cited page was sorted by who published it.
The first table shows the count behind each percentage. Source rows count cited pages, or answers that cited at least one page. The buyer rows use the 210 buyer answers each for the 42 vendors whose site says who they sell to.
I wrote the scoring rules down before the first run. They’re at the bottom of this piece.
Checking the answers turned up eight vendors whose own site names a buyer my directory profile had missed. Those answers were right and my profiles were wrong, so they’re scored right. I’ve updated the directory where the vendor’s page covers the product it profiles.
The limits
- This is two questions on two assistants, run from the UK on one day. Different wording may get different answers.
- It measures the pages each assistant cited, not everything it read.
- Scoring takes judgment. I did it on my own. The rules are below so anyone can check a call they disagree with.
- Nothing here judges the products. It’s only what the assistants said about them.
The scoring rules
Answer key. Each vendor’s Treasury Brief profile on the day of the pull. Where the vendor’s own site said more than the profile, the site won.
Before anything else. An answer about a different company, or about something else with the same name, counts as “different company”. An answer that describes two companies without picking one counts as “undecided”. Neither is scored further.
“What is it”. Scored on the opening description only, before any feature list.
- Right: it names the product category the vendor sells, or describes the company the way its own homepage does.
- Vague: nothing false, but no category a buyer could shortlist on.
- Wrong: it calls the vendor something it isn’t.
“Who is it for”. Scored on the first company size the answer names.
- Startups, small businesses and SMEs count as small. Mid-sized and mid-market count as mid-market. Large companies, enterprises and multinationals count as enterprise.
- Right: every size named is one the vendor’s site claims.
- Wrong buyer: it names a size the vendor’s site doesn’t claim.
- No size given: it describes the buyer by need only, like “companies with many bank accounts”.
“Not a good fit for”. An answer rules out a buyer only when it says a size group the vendor sells to, as a whole, is a poor fit. Only the extreme end (“very small startups”, “massive multinationals”) doesn’t count. Neither does a size tied to a need (“a small business with simple finances”).
Sources. Each cited page counts once per answer. It’s sorted by who published it: the vendor (including its parent company), review and list sites, competitors, consultants and partners, news and press, or other.
From 5 runs to one answer. A vendor’s usual answer is the most common code across its 5 runs. A tie goes to the worse code.
If something here is unclear, reach out to alex@alexcan.xyz.