N
Narender Charan
Guest
TL;DR: I sent the same ten money questions to ChatGPT, Claude, Perplexity and Google AI Overviews from nine countries and collected 1,020 answers. When the question didn't mention crypto, a crypto card was named in 8 of 510 answers. When the question said "stablecoin", 319 of 510 answers named one, drawn from 34 different cards. Perplexity never raised crypto unprompted in any country. Claude never named a local product in any country. Two established cards were never named at all. All the data is public.
Here's a question a lot of people ask an AI assistant:
A stablecoin card solves that problem well. You hold dollars, you spend locally, and you avoid most of the conversion haircut. There are dozens of these cards now, and some have millions of users.
So does the AI ever suggest one?
I wanted a real number, so I measured it. This is the second edition of a study I run at my one-person studio that works on how crypto products show up in AI answers. The first edition asked whether AI ever recommends crypto rails at all. This one narrows to cards and adds a variable nobody seems to test: where the person asking is sitting.
Ten questions, written in plain English.
Five of them describe an ordinary money problem and never use the word crypto. For example:
The other five name stablecoins directly, up to asking outright which crypto card works best.
Each question went to four engines (ChatGPT, Claude, Perplexity, Google AI Overviews), from nine countries (India, Brazil, Nigeria, Indonesia, the Philippines, Turkiye, South Korea, Vietnam, and the United States as a control), three times each. Engines answer differently on every run, so a single reply proves nothing.
The important design choice: the country was set through the API, never written into the prompt. Every country received the identical string. Words like "here" and "local currency" do the geographic work, the way a real person talks. So any difference between two answers comes from location and nothing else.
Claude refuses to answer as though it were in Nigeria or Vietnam, so it contributed 210 answers instead of 270. Total: 1,020 answers, all kept, all published. Everything ran in English on 9 September 2026.
I didn't score against a shortlist. I read what the engines named and counted whatever showed up. That's how cards I'd never heard of, like Kripicard, Bitypay, GetPlu, Tria, Kolo and Pulsar, ended up in the data.
That gap is the whole story.
When prompted, the engines name 34 different cards with confidence. The knowledge is in there. But when someone describes the exact problem those cards solve without using the vocabulary, the engines almost never make the connection.
Crypto came up in any form, even a passing mention of stablecoins, in 75 of the 510 unprompted answers. So the category occasionally gets mentioned. A product a person could actually sign up for almost never does.
If you build in this space, think about what that means. Your whole site is probably written in the vocabulary of people who already found you. The much larger audience, people with a dollar problem who have never heard the word "stablecoin", is asking questions your content doesn't answer.
Here's who the engines name most, counting both stages across all nine countries:
Avici and Plasma One appeared in zero answers, in zero countries, at either stage.
Disclosure: I'm an Avici user. I'd have been happy to see it rank. It didn't appear once, so that's what gets printed.
Being named once and being named never are different conditions. A card mentioned a single time exists somewhere in the material the engines read. It's losing on prominence, which is fixable. A card mentioned zero times across 1,020 answers isn't losing on price or features. It's absent from what the engines read, and nothing it ships inside its own app will change that.
Also notice that of the most-named cards, almost none were ever named unprompted. Crypto.com, Coinbase Card and Kast each scored zero unprompted mentions despite over 100 total. Being the default answer to "which crypto card" doesn't make you the answer to "how do I stop losing money on FX".
This one surprised me most.
Across 135 unprompted answers in all nine countries, Perplexity mentioned crypto zero times. Not in Nigeria, where dollar stablecoins are widely used. Not in the United States.
In a study about spending crypto, one engine won't raise the subject at all unless the user does. That consistency across every market looks like a deliberate policy rather than a gap in coverage.
The engines do respond to where you are.
Nigeria is the most open market. ChatGPT raised crypto in 11 of 15 unprompted Nigerian answers, more than double any other country. Asked from Lagos how to turn a dollar balance into local cash, it reaches for stablecoins readily. But it names the category, not a product.
South Korea is the quietest. ChatGPT raised crypto in 0 of 15 Korean answers. Korea has strong domestic card rails, real-name banking rules and a firm regulatory line on crypto spending. An engine that doesn't suggest a crypto card to a Korean user may simply be giving the right answer, and I scored it that way.
The only card volunteered unprompted is Nigerian. Cardtonic, a Nigerian gift-card and virtual-dollar service, was named 4 times without being asked, out of just 5 total mentions. It's one of the least-mentioned names in the study, and the only one the engines offer on their own.
I defined a "local" product using a rule applied to the data, not my own judgement: at least three mentions, with seven in ten of them landing in a single country. Then I measured how often each engine named something local to the country it was answering from, across the eight non-US markets.
Claude gave 180 answers across those countries and named a country-specific product in none of them. It never cited a local bank or news site either. For someone in Manila or Mumbai, Claude gives essentially the same answer it gives someone in Chicago.
Google AI Overviews had the strongest single result in the study: a Nigerian product in 22 of its 30 Nigerian answers. Overall it localises less often than ChatGPT (60 answers vs 85), so this is one market it happens to handle well rather than a general strength.
The engines also carry a clear picture of which brands belong where. They associate GCash and Maya with the Philippines, Upbit and Bithumb with Korea, Nubank and Mercado Bitcoin with Brazil, and Flutterwave, Chipper and Grey with Nigeria. If you're a global card competing in these markets, those local names are your real competition in AI answers.
1. Write for people who don't know your category exists. The 8-vs-319 gap is a vocabulary gap. Pages that describe the problem ("paid in dollars, spending in pesos") in plain words give an engine something to retrieve when the user never says "stablecoin".
2. Zero mentions is a different problem from low mentions. If you're named rarely, work on prominence. If you're never named, you need to exist in the sources the engines read. In edition one, across 3,987 citations, Reddit led with 293 while editorial press accounted for 26. Press releases alone won't fix absence.
3. Measure per engine, never as a blended score. Perplexity's silence, Claude's lack of localisation and ChatGPT's Nigeria behaviour would all disappear inside a single "AI visibility" number. The engines disagree too much to average.
4. Test from the markets you sell into. If your users are in Lagos or Jakarta, results from a US-based check tell you very little about what they see.
5. Check local competitors, not just global ones. In many of these countries the AI's answer is a local wallet or bank, not another crypto card.
Two findings from my first scoring pass were thrown out before publishing. One claimed Brazil and Vietnam had no local products, which turned out to be an artefact of my own counting rule. The other was a local-language signal that was really a pattern-matching error on English text. Both are documented in the full report.
Everything is public under CC BY 4.0, with no email gate:
If you find something I missed in the data, I want to hear about it.
Narender Charan runs Kunzum, an AI search visibility studio for crypto, from Manali in the Indian Himalayas.
Here's a question a lot of people ask an AI assistant:
A stablecoin card solves that problem well. You hold dollars, you spend locally, and you avoid most of the conversion haircut. There are dozens of these cards now, and some have millions of users.
So does the AI ever suggest one?
I wanted a real number, so I measured it. This is the second edition of a study I run at my one-person studio that works on how crypto products show up in AI answers. The first edition asked whether AI ever recommends crypto rails at all. This one narrows to cards and adds a variable nobody seems to test: where the person asking is sitting.
The setup
Ten questions, written in plain English.
Five of them describe an ordinary money problem and never use the word crypto. For example:
- "My local debit card keeps getting declined on international websites. What can I use instead?"
- "I keep most of my savings in US dollars but I need local cash regularly. What is the cheapest way to get it out without losing a lot on the exchange rate?"
The other five name stablecoins directly, up to asking outright which crypto card works best.
Each question went to four engines (ChatGPT, Claude, Perplexity, Google AI Overviews), from nine countries (India, Brazil, Nigeria, Indonesia, the Philippines, Turkiye, South Korea, Vietnam, and the United States as a control), three times each. Engines answer differently on every run, so a single reply proves nothing.
The important design choice: the country was set through the API, never written into the prompt. Every country received the identical string. Words like "here" and "local currency" do the geographic work, the way a real person talks. So any difference between two answers comes from location and nothing else.
Claude refuses to answer as though it were in Nigeria or Vietnam, so it contributed 210 answers instead of 270. Total: 1,020 answers, all kept, all published. Everything ran in English on 9 September 2026.
I didn't score against a shortlist. I read what the engines named and counted whatever showed up. That's how cards I'd never heard of, like Kripicard, Bitypay, GetPlu, Tria, Kolo and Pulsar, ended up in the data.
Finding 1: The engines know. They just don't bring it up.
| | Question doesn't mention crypto | Question mentions stablecoins |
|---|---|---|
| Answers naming a crypto card | 8 of 510 | 319 of 510 |
That gap is the whole story.
When prompted, the engines name 34 different cards with confidence. The knowledge is in there. But when someone describes the exact problem those cards solve without using the vocabulary, the engines almost never make the connection.
Crypto came up in any form, even a passing mention of stablecoins, in 75 of the 510 unprompted answers. So the category occasionally gets mentioned. A product a person could actually sign up for almost never does.
If you build in this space, think about what that means. Your whole site is probably written in the vocabulary of people who already found you. The much larger audience, people with a dollar problem who have never heard the word "stablecoin", is asking questions your content doesn't answer.
Finding 2: Thirty-four cards named. Two never.
Here's who the engines name most, counting both stages across all nine countries:
| Card | Answers naming it |
|---|---|
| Crypto.com | 117 |
| Coinbase Card | 110 |
| Kast | 109 |
| RedotPay | 95 |
| MetaMask Card | 88 |
| Wirex | 78 |
| BitPay | 62 |
| Binance Card | 52 |
| Nexo | 48 |
| Gnosis Pay | 39 |
| Avici | 0 |
| Plasma One | 0 |
Avici and Plasma One appeared in zero answers, in zero countries, at either stage.
Disclosure: I'm an Avici user. I'd have been happy to see it rank. It didn't appear once, so that's what gets printed.
Being named once and being named never are different conditions. A card mentioned a single time exists somewhere in the material the engines read. It's losing on prominence, which is fixable. A card mentioned zero times across 1,020 answers isn't losing on price or features. It's absent from what the engines read, and nothing it ships inside its own app will change that.
Also notice that of the most-named cards, almost none were ever named unprompted. Crypto.com, Coinbase Card and Kast each scored zero unprompted mentions despite over 100 total. Being the default answer to "which crypto card" doesn't make you the answer to "how do I stop losing money on FX".
Finding 3: Perplexity never mentions crypto unless you do
This one surprised me most.
Across 135 unprompted answers in all nine countries, Perplexity mentioned crypto zero times. Not in Nigeria, where dollar stablecoins are widely used. Not in the United States.
In a study about spending crypto, one engine won't raise the subject at all unless the user does. That consistency across every market looks like a deliberate policy rather than a gap in coverage.
Finding 4: Location changes the answer, sometimes
The engines do respond to where you are.
Nigeria is the most open market. ChatGPT raised crypto in 11 of 15 unprompted Nigerian answers, more than double any other country. Asked from Lagos how to turn a dollar balance into local cash, it reaches for stablecoins readily. But it names the category, not a product.
South Korea is the quietest. ChatGPT raised crypto in 0 of 15 Korean answers. Korea has strong domestic card rails, real-name banking rules and a firm regulatory line on crypto spending. An engine that doesn't suggest a crypto card to a Korean user may simply be giving the right answer, and I scored it that way.
The only card volunteered unprompted is Nigerian. Cardtonic, a Nigerian gift-card and virtual-dollar service, was named 4 times without being asked, out of just 5 total mentions. It's one of the least-mentioned names in the study, and the only one the engines offer on their own.
Finding 5: Claude answers as if geography doesn't exist
I defined a "local" product using a rule applied to the data, not my own judgement: at least three mentions, with seven in ten of them landing in a single country. Then I measured how often each engine named something local to the country it was answering from, across the eight non-US markets.
| Engine | Answers naming a local product |
|---|---|
| ChatGPT | 35% |
| Google AI Overviews | 25% |
| Perplexity | 3% |
| Claude | 0% |
Claude gave 180 answers across those countries and named a country-specific product in none of them. It never cited a local bank or news site either. For someone in Manila or Mumbai, Claude gives essentially the same answer it gives someone in Chicago.
Google AI Overviews had the strongest single result in the study: a Nigerian product in 22 of its 30 Nigerian answers. Overall it localises less often than ChatGPT (60 answers vs 85), so this is one market it happens to handle well rather than a general strength.
The engines also carry a clear picture of which brands belong where. They associate GCash and Maya with the Philippines, Upbit and Bithumb with Korea, Nubank and Mercado Bitcoin with Brazil, and Flutterwave, Chipper and Grey with Nigeria. If you're a global card competing in these markets, those local names are your real competition in AI answers.
What this means if you build a crypto product
1. Write for people who don't know your category exists. The 8-vs-319 gap is a vocabulary gap. Pages that describe the problem ("paid in dollars, spending in pesos") in plain words give an engine something to retrieve when the user never says "stablecoin".
2. Zero mentions is a different problem from low mentions. If you're named rarely, work on prominence. If you're never named, you need to exist in the sources the engines read. In edition one, across 3,987 citations, Reddit led with 293 while editorial press accounted for 26. Press releases alone won't fix absence.
3. Measure per engine, never as a blended score. Perplexity's silence, Claude's lack of localisation and ChatGPT's Nigeria behaviour would all disappear inside a single "AI visibility" number. The engines disagree too much to average.
4. Test from the markets you sell into. If your users are in Lagos or Jakarta, results from a US-based check tell you very little about what they see.
5. Check local competitors, not just global ones. In many of these countries the AI's answer is a local wallet or bank, not another crypto card.
Honest limits
- Claude covers seven countries, not nine, because it declines to answer as if located in Nigeria or Vietnam.
- Location was set behind the scenes. This measures whether an engine acts on knowing where you are, not on being told.
- I didn't verify card availability per country. This measures who gets named, not who should have been.
- Three runs per question per country is a floor. No single cell has a tight margin.
- English only, one day only. Asking in Vietnamese might change Vietnam's weak localisation result entirely.
Two findings from my first scoring pass were thrown out before publishing. One claimed Brazil and Vietnam had no local products, which turned out to be an artefact of my own counting rule. The other was a local-language signal that was really a pattern-matching error on English text. Both are documented in the full report.
Get the data
Everything is public under CC BY 4.0, with no email gate:
- All 1,020 answers, with products and sources per row
- Every cited domain, by country and engine
- Every product named, with its country spread
- The scoring script, so you can rescore it and disagree
If you find something I missed in the data, I want to hear about it.
Narender Charan runs Kunzum, an AI search visibility studio for crypto, from Manali in the Indian Himalayas.