Research / Edition one / September 2026
Ten ordinary money problems that stablecoin rails genuinely solve, described in plain language to four AI engines. Then asked again, with crypto named.
The finding
Ask an engine, in ordinary words, how to get paid by a client in another country, and it will raise crypto surprisingly often. Across 360 unprompted answers it happened 35% of the time, and ChatGPT did it in 61%. We expected under 20%. That prediction failed badly, and it is published below with the rest.
Naming something is different. A person who needs to invoice a client on Thursday cannot act on the word “stablecoin.” They need a product. Across those same 360 answers, 1.7% named one — six answers.
Add the word crypto to the question and the names arrive at once: Coinbase, Circle, KAST, Wirex, Bitwage. The engines are not refusing the category. They are declining to name anyone inside it unless you already know to ask.
If your product is the answer and the engine will not say your name, you are not losing on merit. You are losing on retrieval.
What we asked
Ten scenarios, each a real problem where stablecoin rails are honestly competitive. A freelancer in Lagos paid by a client in Berlin. A designer in Buenos Aires watching her savings lose value. A developer in Hanoi who cannot open a US account. An agency paying eight contractors in five countries. No trading, no yield, no speculation. If crypto was not genuinely the better answer, the scenario did not go in.
Stage one describes the problem the way a person would, with no hint crypto exists. Stage two asks the same thing and says the word. Stage two is the control: if the engine answers well when asked directly, the knowledge existed, and the gap between the two is what it knows but does not volunteer.
Each scenario is written three ways — neutral, region-specific and frustrated — because real people write all three, and phrasing changes what comes back. Thirty stage-one prompts, ten stage-two, run three times each against ChatGPT, Claude, Perplexity and Google AI Overviews.
The numbers
| Engine | Stage 1 raised crypto | Stage 1 named a product | Stage 2 raised crypto | Gap |
|---|---|---|---|---|
| ChatGPT | 61.1% | 1.1% | 100% | +38.9 |
| Claude | 28.9% | 0.0% | 100% | +71.1 |
| Google AI Overviews | 34.4% | 5.6% | 100% | +65.6 |
| Perplexity | 16.7% | 0.0% | 100% | +83.3 |
| All | 35.3% | 1.7% | 100% | — |
ChatGPT raises crypto nearly four times as often as Perplexity on identical prompts. Across the forty prompts, thirteen were total splits — one engine raised crypto in all three runs while another raised it in none. The engines also disagreed with themselves: three runs of the same prompt gave different answers in 20% of cases.
There is no single “what AI says about crypto.” There are four answers, and they do not agree.
M4 · what gets recommended instead
The four engines named 90 distinct products across the 480 responses. Wise leads at 246, ahead of PayPal at 156, Payoneer at 139 and Stripe at 127. The first crypto-native name is Circle, at 14.
The overall ranking hides how much the answer moves with the problem. Four different products lead the ten scenarios: Mercury owns the de-banked business, M-Pesa owns NGO disbursement, PayPal owns the declined-card question.
The full ranking and every scenario →M5 · caution asymmetry
A crypto recommendation is 2.8 times more likely to arrive with a caution attached than a traditional one: 46% against 16%, and the ratio holds across all four engines.
All 375 unique warning sentences were then read and classified by hand. 82% of the crypto ones are accurate, against 73% for traditional products. The caution is disproportionate in volume and proportionate in content.
That page also carries the counter-case: the one scenario where the engines are right to steer a reader away from crypto, and correctly say that stablecoin rails will not fix a compliance-driven de-banking.
Every warning, classified →M6 · where the answers come from
The corpus logged 3,987 citations across 849 domains. reddit.com leads at 293. Vendor pages account for 3,409 citations and editorial press for 26.
The four engines also read different internets. Perplexity alone cites Reddit 222 times. And enabling web search only permits retrieval: ChatGPT ran a search on 54% of its answers, Perplexity on 100%.
Every domain, by engine and type →Registered before the run
These were written down before a single call was made, which is the only thing separating research from an asset assembled afterwards to fit a conclusion. The failures are printed with the same weight as the wins.
It was 35.3%, and 61% for ChatGPT. Only Perplexity came in under, at 16.7%.
100% surfacing on all four engines, 45% naming a specific product.
Wise tops 5 of 10 scenarios and Payoneer is top-three in 5, but merchant, de-banking and NGO scenarios each have a different leader.
2.8× overall, consistent from 2.1× to 3.3×. At the low end of “several”.
Runs disagreed within an engine in 20% of cells; 13 of 40 prompts were total splits between engines.
Coinbase leads at 41 mentions and has both. No clean separation.
Limitations
Every prompt was asked from a US location. Holding it constant isolates phrasing as the variable, which is what the three phrasings are for. But these engines localise, and this study did not let them.
Half of ChatGPT’s answers never touched the web. Enabling web search permits retrieval, it does not force it. ChatGPT searched in 54% of calls, Claude 65%, Perplexity 100%. The source numbers should be read with that in mind.
Per-scenario slices are thin. Nine runs at stage one and three at stage two for a single engine. At three runs the only possible rates are 0, 33, 67 and 100%. Scenario-level claims in this report aggregate across engines.
Claude declined one scenario entirely. All three runs of the “PayPal and Stripe don’t operate in my country” prompt came back asking which country, recommending nothing. Counted as declined, not as an absence of crypto.
Temperature was 0.94, the API default rather than anything a consumer app uses. That inflates the run-to-run variance the fifth hypothesis is about.
If you build one of these products
This is the useful conclusion, and it is more encouraging than the one we expected to find. The engines are not hostile to the category. They raise it unprompted a third of the time, they describe it accurately, and their warnings are fair. When asked directly they produce competent, specific answers naming real products.
So the thing standing between your product and the answer is not a model that dislikes crypto. It is that when someone describes your exact use case in their own words, nothing in the retrieved material connects that description to your name. Six times out of 360, something did.
That is a fixable problem, and this study says where to spend. Vendor documentation and forum threads are what got cited, at 3,409 and 493 respectively, against 26 for editorial press. The engines pull from your own pages and from what people say about you in places like Reddit. Not from your announcement.
Kunzum does this work for crypto protocols — running buyers’ real questions across the engines, then changing what comes back. Book twenty minutes if that is your problem, or read the method first.
The data
The full corpus behind this report is free to download, with no email form in front of it. It is published under Creative Commons Attribution 4.0, so you can rescore it and publish a different answer.
Optional