Three AI engines, eight Dutch money questions

240 conversations across ChatGPT, Gemini and Claude, leading to 720 answers. Completed on 10 October 2026.

People don't search for financial advice anymore, they ask. Whether it is a couple in Utrecht wondering if the €150,000 they clear on their house can also pay for a new car, or an engineer a year into a job in Eindhoven working out whether she has to file a Dutch tax return at all.

How many people do this depends on who is counting. A bunq survey of 1,000 Dutch adults, published in March 2026, put it at 55%. The AFM's Consumentenmonitor 2026, published in September, put it at about one in seven over the past year. Either way, that is a lot of conversations in which a mortgage adviser, a broker or a tax firm is named, or is not.

So we ran the conversations, on three AI engines: ChatGPT, Gemini and Claude. We took eight questions from published research on what people in the Netherlands ask about money, four in Dutch and four in English, and asked each one ten times on each engine. Every answer was followed by two replies from a simulated buyer: first their personal details, then, in seven of the ten runs, the question every buyer gets to eventually, "who should I go to?". In the other three, the buyer only asked "what's my first step?", to see what a buyer who never asks for names is shown.

The same question, put to three engines, got three different kinds of answer.

Five findings

1. A published price gets a firm named, on two engines of three

When the buyer asked what advice would cost, ChatGPT and Claude went looking for pages that state a price, and named the firms behind them. ChatGPT reached 27 named firms through a price or tariff page and quoted the published figure next to the name. Claude reached 25.

Gemini reached 11. It usually gave a price range with no page behind it, and it searched in only 15 of the 56 third answers, the turn where it lists firms and prices.

A page that states what a service costs, in plain figures, is something two of the three engines can find and quote when a buyer asks that exact question.

2. An expat and a Dutch resident are shown two different markets

On every engine, English conversations named brands earlier, and the two languages shared few names.

ChatGPT did not name a single company in any Dutch first answer. A firm that publishes only in Dutch is close to absent for a buyer who asks in English, and the other way round.

3. The same buyer gets a different figure each time, or none

Each row is one buyer with identical details. The ranges are the lowest and highest figure given across ten runs.

ChatGPT explained the rules and declined to put a number on the buyer's own case, with one exception: the expat couple's mortgage, where its range was the tightest of the three. Gemini and Claude gave numbers every time, and the numbers moved. For the expat couple, one Claude run put their limit at €400,000 to €450,000 and another at €480,000 to €550,000, so the €475,000 flat they were looking at was out of reach in one conversation and within reach in another.

4. ChatGPT waits to be asked, Gemini volunteers, Claude sits in between

We counted the conversations in which an engine put a firm forward as an option before the buyer asked for names.

ChatGPT explains the rules, points to the Belastingdienst and the AFM, and keeps company names back until the buyer asks who to go to. Gemini names firms in its first answer more often than not. Claude does so in about a quarter of conversations, mostly on the two investing questions.

The difference matters most for the buyer who never asks. In the three runs per question where the buyer only asked "what's my first step?", ChatGPT put a firm forward in 1 of 24 conversations. Gemini did so in 16 of 24, with 4.5 firms on average. Claude did so in 6 of 24, almost all of them about investing.

5. Only 27 brands appear on all three engines

Across the three engines, 252 different commercial brands were named. Twenty-seven of them appeared on all three. ChatGPT named 56 brands in total, Gemini 142 and Claude 153, and on every engine about 40% of the brands appeared in a single conversation and never again.

The engines differ in how settled their shortlist is. Gemini repeats its leading names: for some questions it named the same nine firms in all seven runs where the buyer asked. ChatGPT's list is short and fairly stable. Claude names about as many brands as Gemini but repeats few of them, and on five of the eight questions no brand appeared in all seven of its runs.

There is one question where all three agree. On the pension question, Brand New Day and Meesman were put forward in every run, on every engine.

And this is not all

Each engine also has its own way of handling trust.

• ChatGPT cites. It searched in every first and second answer and linked a source in 96% of all answers. The Belastingdienst's site was cited in 68% of its answers, and its four most-cited sites are all public bodies.
• Gemini asserts. It linked sources in 63% of answers, named firms and prices from memory, and never once said a price or a name was unverified. It also told buyers that paid help was not worth it in 15 of 80 second answers, mostly on box 3 and the first tax return.
• Claude hedges. In 43 of the 56 conversations where the buyer asked for names, it said it had not checked the prices or quality of the firms it named. In 11 it declined to name individual advisers at all, mostly on mortgages and pensions, saying it cannot judge who is good from a search result.

The AFM register is a Claude and ChatGPT habit: they pointed buyers to it in 23 and 13 of those 56 third answers. Gemini did not mention it once.

Which brands each engine named

The table shows the three firms each engine put forward most often as an option for each question, across all ten conversations per question. The number is how many of those ten conversations named the firm. Where several firms tie, all of them are listed.

Being put forward and being cited are mostly different lists. Claude and Gemini also read firms' own pages for facts without suggesting the firm. On Claude these were explainer pages from HomeFinance, Knab, Van Lanschot Kempen and Financer. On Gemini, KRC van Elderen's pages were cited in 22 conversations, mostly on the expat tax and investing questions, without the firm being named once. ChatGPT drew its facts almost entirely from public bodies, so its cited firms are close to none. The exceptions run the other way: on Gemini, Viisi was both cited and put forward on the expat mortgage question, and TaxSavers on the expat tax questions.

Three patterns stand out.

Low-cost providers lead wherever investing comes up. All three engines steer towards a global index fund, bought through a low-cost broker or a ready-made index plan. Meesman and Brand New Day lead the pension question on every engine, and Vanguard, DEGIRO and Interactive Brokers recur on investing.

Gemini brings the big banks in. ING, Rabobank and ABN AMRO appear in every Gemini run of both investing questions. On ChatGPT and Claude the large banks appear less often, and on Claude mostly as a cited explainer page.

Specialists win on specialist questions. On the expat tax questions the names are small firms that serve internationals, such as TaxSavers, Blue Umbrella, JC Suurmond and Broadstreet. In these lists the large accounting firms appear only on Claude, for the 2027 question.

Where the answers link. ChatGPT linked a source in 96% of its answers, Claude in 81% and Gemini in 63%. ChatGPT's sources are mostly public: belastingdienst.nl, afm.nl and rijksoverheid.nl lead. Gemini's most-cited sites after the Belastingdienst are commercial: taxsavers.nl, krcvanelderen.nl, abnamro.nl and rabobank.nl. Claude leans on intermediaries, such as iamexpat.nl, the Consumentenbond, bank explainer pages and comparison sites.

Top brands by presence score

Counting mentions alone treats a firm an engine volunteers the same as one it produces only under questioning, and a firm put forward as an option the same as one whose web page was merely cited. So we scored each brand per conversation, counting only the times it was put forward as an option:

• 2 points when the engine named the firm before the buyer asked for names.
• 1 point when it named the firm only after the buyer asked.
• 0 points when the firm did not appear.

The score is the total as a percentage of the maximum possible, which is 2 points in each of the 80 conversations an engine had. A score of 100 would mean a firm was volunteered in every conversation on every question. Being volunteered counts double because it reaches every buyer, including the ones who never ask "who?".

Five firms are in the top ten on all three engines: Meesman, Brand New Day, ABN AMRO, Vanguard and Interactive Brokers.

Two things to keep in mind when reading it.

The scales differ by engine, and that is part of the finding. Gemini's scores are several times the others' because it names firms early: a firm Gemini volunteers in the first answer scores 2 in every conversation, where a firm ChatGPT names only when asked scores 1. A ChatGPT score of 9.4 and a Gemini score of 45 can both mean "a leading name on this engine".

A score rewards breadth. A bank can come up in a mortgage, a pension and an investing conversation, while an expat tax firm can only come up on one or two of the eight questions. Claude's top ten is mostly brokers and funds because its two investing questions are where it names firms early and repeats them.

The role filter changes the ranking. When every mention counts, including pages cited as a source, Rabobank leads Gemini and Knab, HomeFinance and Raisin enter Claude's top ten for their explainer pages. KRC van Elderen's pages were cited in 22 Gemini conversations without the firm being named once.

How we ran it

1. Questions from research. Eight starting questions, each tied to a published source on what people in the Netherlands ask about money. Each states a problem in the buyer's own words and names no brand.
2. Three engines, a fresh session each time. Every conversation started with no memory of any other. Web search was available and the engine decided whether to use it. Each engine was asked to act as a consumer chat assistant and link its sources, and was not told it was part of a study.
3. A simulated buyer. Each question had one fixed buyer profile, used in every run on every engine. The second message gave personal details and asked "what would you do in my case?". In runs 1 to 7 the third message asked who to go to and what it would cost. In runs 8 to 10 it asked only "what's my first step?".
4. Ten runs of three answers. 8 questions, 10 runs, 3 engines: 240 conversations and 720 answers.
5. The reading. Every answer was read for named organisations, each tagged by type, by the answer it first appeared in, and by role: put forward as an option, listed as an example, cited as a source, or cautioned against. A 10% sample per engine was checked by hand, with no invented names and no missed commercial names.

The eight questions

Comparing insurance, the most common use in the Multiscope survey at 22%, was not part of this study.

Where does your brand stand?

The same buyer, asking the same question, can meet a firm in the first answer on one engine and not at all on another. Which of those a potential customer sees depends on the engine they open, the language they ask in and whether they ever get round to asking "who should I go to?".

The more useful version of this exercise starts from your side: your buyers, the questions they would put to an assistant, your markets and languages, and the competitors you name.

If you would like to know where your brand stands, we are happy to talk it through. In a 20-minute call we look at how a potential customer would describe the problem you solve, what ChatGPT, Gemini and Claude say when they do, and whether optimising for the AI buyer journey is worth your attention right now. Sometimes the answer is that it can wait, and we will say so.

The call is free and you leave with a first read on your position either way.

Book a 20-minute call