Research / The Crypto Card Report / Edition two
We asked four AI engines the same ten questions from nine countries. Nothing changed but where the question came from.
The finding
8 of 510
Five of our ten questions describe an ordinary money problem and never mention crypto. A crypto card genuinely answers all five. Across 510 of those answers, from nine countries, one was named 8 times.
Ask the same thing while saying the word stablecoin, and 319 of 510 answers name a card, from a field of 34 different products. The engines know. They just never bring it up.
We did not score against a shortlist. We read what the engines said and counted whatever they named, which is how Kripicard, Bitypay, GetPlu, Tria, Kolo and Pulsar ended up in this study.
Being named once and being named never are different conditions. A card mentioned a single time is somewhere in the material these engines read, and is losing on prominence. A card mentioned never is not in that material at all, and nothing it does inside its own app will change that.
| Card | Answers naming it | Of those, unprompted | Countries |
|---|---|---|---|
| Crypto.com | 117 | 0 | 9 |
| Coinbase Card | 110 | 0 | 9 |
| Kast | 109 | 0 | 9 |
| RedotPay | 95 | 0 | 9 |
| MetaMask Card | 88 | 0 | 9 |
| Wirex | 78 | 0 | 9 |
| BitPay | 62 | 1 | 9 |
| Binance Card | 52 | 0 | 9 |
| Nexo | 48 | 0 | 9 |
| Gnosis Pay | 39 | 0 | 8 |
| Bybit Card | 37 | 1 | 9 |
| Ether.fi | 34 | 0 | 9 |
| Oobit | 33 | 0 | 9 |
| Bleap | 32 | 1 | 9 |
| Coinbase Commerce | 25 | 0 | 7 |
| Bitrefill | 22 | 0 | 8 |
| MoonPay | 21 | 0 | 9 |
| Kripicard | 21 | 0 | 9 |
| BVNK | 18 | 0 | 7 |
| CoinGate | 15 | 0 | 6 |
| Bitget Wallet | 14 | 0 | 7 |
| Bitypay | 10 | 0 | 6 |
| Rain | 8 | 0 | 4 |
| Kolo | 8 | 0 | 6 |
| GetPlu | 7 | 1 | 4 |
| Pulsar | 7 | 0 | 2 |
| Tria | 7 | 0 | 4 |
| Holyheld | 5 | 0 | 3 |
| Cardtonic | 5 | 4 | 1 |
| Sphere | 3 | 0 | 3 |
| Yellow Card | 3 | 0 | 2 |
| Bitnob | 2 | 0 | 1 |
| Bit2Me | 1 | 0 | 1 |
| Jupiter Global | 1 | 0 | 1 |
| Avici | 0 | 0 | 0 |
| Plasma One | 0 | 0 | 0 |
Aliases are merged: Kast appears in the corpus as both KAST and Kast Card, and Wirex Card never appeared without Wirex, so counting both would inflate every total.
A card was named 8 times in 510 answers. Crypto came up in any form, even a passing mention of stablecoins, 75 times.
In 135 answers across all nine countries, Perplexity mentioned crypto 0 times. Not in Nigeria, not in the United States. In a study about spending crypto, one engine declines to mention the subject unless you do. That looks like a decision rather than a gap.
ChatGPT raised crypto in 11 of 15 Nigerian answers, more than twice any other country. Asked from Lagos how to turn a dollar balance into local cash, it reaches for stablecoins readily. The engines do respond to where you are. They respond by mentioning the category, never by naming a product.
ChatGPT raised crypto in 0 of 15 Korean answers. Korea has strong domestic card rails, real-name banking rules and a hard regulatory line on crypto spending. An engine that does not suggest a crypto card to a Korean may simply be right, and this report treats that as a fair answer rather than a failure.
Cardtonic, a Nigerian gift-card and virtual-dollar service, was named 4 times without being asked, out of 5 appearances in total. It is the only card in the study the engines offer unprompted, and one of the least-mentioned names in it.
A product counts as local here if seven in ten of its mentions across the whole study land in one country. We applied that rule to the data rather than deciding by hand which names felt regional. These rates cover the eight countries outside the United States, which is a control rather than a market being served.
Across the eight countries outside the United States control, Claude gave 180 answers and named a product specific to the country it was answering in none of them. It never cited a local news site or bank either. It also refuses to answer as though it were in Nigeria or Vietnam at all. For a reader in Lagos, Claude is the same product it is for a reader in Chicago.
Google AI Overviews named a Nigerian product in 22 of its 30 Nigerian answers, the highest single engine-and-country result anywhere in the study. Across all eight countries it localises less often than ChatGPT, 60 answers against 85, so this is one country it is unusually good at rather than a general strength.
| Country | Adoption rank | Names the study found are specific to it |
|---|---|---|
| India | 1 | CoinDCX, Cryptomus, Mudrex, Niyo, WazirX, ZebPay |
| Brazil | 5 | Bitomat, C6, Conta Global, Foxbit, Mercado Bitcoin, Nomad, Nubank, TED |
| Nigeria | 6 | Busha, Cardtonic, Chipper, Cleva, Flutterwave, GTBank, Geegpay, GrabrFi, Grey |
| Indonesia | 7 | BCA, CIMB Niaga, GoPay, Indodax, Mandiri, OVO, Pintu, Tokocrypto |
| Philippines | 9 | BDO, BPI, GCash, GoTyme, InstaPay, Maya, PDAX, PayMaya |
| Turkiye | 14 | BtcTurk, Papara, Paribu |
| South Korea | 15 | Bithumb, Coinone, DaWinKS, Kaia, Korbit, Upbit |
| Vietnam | 4 | BitcoinVN, MoMo, NAPAS, Techcombank, Vietcombank, ZaloPay |
| United States | 2 | Coinsbee, Gemini |
Countries are the top crypto-adoption markets by the Chainalysis 2025 index, excluding mature markets. The United States is included only as a yardstick.
Every prompt was sent as the same string to every country. The location was set through the API rather than written into the question, so no prompt names a country. The words “here” and “local currency” do that work, which is how a real person asks.
That means any difference between two answers is down to where the person asking was sitting, and nothing else.
The other five say stablecoin out loud, down to asking outright which crypto card works best. All ten are in prompts.json, and the full method is in the PDF.
The data
The whole corpus is free to download, with no email form in front of it, under Creative Commons Attribution 4.0. Rescore it and publish a different answer.
Optional
Claude does not cover two of the nine countries. It refuses to answer as though it were in Nigeria or Vietnam, so its figures rest on seven countries while the others rest on nine.
We set the location rather than saying it. That isolates geography cleanly, but it measures whether an engine acts on knowing where you are, not whether it acts on being told.
We did not check whether products are available. This says who gets named. It does not claim a named card is available to you, or that an unnamed one should have been.
Three tries is a floor. These engines answer differently each time. Three repetitions per question per country means nothing rests on a single reply, and not that any individual cell has a tight margin.
English only, one day only. Asked in English on 9 September 2026. Vietnam was the weakest country for local answers, and asking in Vietnamese might change that entirely.
Two claims our first scoring pass produced were discarded before publication: that Brazil and Vietnam had no local products at all, which was an artefact of our own counting rule, and an apparent local-language signal that turned out to be a pattern-matching error on English text. Both are recorded in the PDF, because a report that shows only the analyses that worked is not showing its method.
Kunzum. The Crypto Card Report, edition two. September 2026.
https://kunzum.xyz/research/which-crypto-cards-ai-recommends/
Edition one asked whether AI ever tells you crypto is the answer →