# Does AI ever tell you crypto is the answer? ## Methodology for the Crypto AI Visibility Index, edition one Version 2.0. Prepared for Kunzum. --- # 1. The question Most crypto research asks which protocol wins. That is trade press, and only crypto people read it. This asks something a general reader cares about: > **When an ordinary person describes a real problem that stablecoin rails > genuinely solve better than anything else, does an AI engine ever tell them?** A freelancer in Lagos paid by a client in Berlin loses days and a chunk of the invoice to correspondent banking. A designer in Buenos Aires watching her savings lose value cannot easily hold dollars. A developer in Karachi cannot open a US account to receive payouts. Products like Avici, KAST, Kosh and Bleap solve these with stablecoin rails behind the interface. The user does not need to know or care that crypto is involved. It simply works, and for many of these people it is the most efficient option that exists. The question is whether the engine that person asks will ever mention it. ## The second question, which is the honest one If the engines do not surface these products, is that wrong? Crypto has earned a large share of its stigma. Scams, hacks, collapses, and a decade of products that lost people money. An engine that hedges when recommending a crypto-adjacent financial product is not obviously malfunctioning. It may be doing its job. So the report does not argue that AI is biased against crypto. **It measures whether the caution is proportionate**, by comparing how the same engine treats a stablecoin card and a traditional fintech card recommended to the same user for the same task. If the answer turns out to be that the caution is warranted, we publish that. A report willing to reach that conclusion is the only kind a journalist will trust. --- # 2. Design: two stages Every scenario is asked twice. **Stage 1, neutral.** The user describes their problem in ordinary language. No crypto vocabulary, no hint that crypto exists. > "I'm a freelancer in Nigeria and my clients in Europe pay me by bank transfer. > It takes a week and the fees are brutal. What are my options?" **Stage 2, prompted.** Same scenario, crypto explicitly raised. > "Same situation. Is there a crypto or stablecoin option that would work better?" Stage 2 is the control. If the engine gives a competent, specific answer at stage 2, the knowledge existed. The gap between stage 2 and stage 1 measures what the engine **knows but does not volunteer**, and that is a far stronger finding than absence alone. --- # 3. Scenarios Ten. Each is a real, common, non-crypto problem where stablecoin rails are genuinely competitive. No speculation, no trading, no yield chasing. If crypto is not honestly the better answer, the scenario does not belong in the study. | # | Scenario | Who has this problem | Why crypto rails genuinely help | |---|---|---|---| | 1 | Freelancer paid from abroad, losing days and fees | Africa, South Asia, LatAm, SE Asia | Settlement in minutes, fees in cents | | 2 | Needs a US account and card without being American | Developers, freelancers, small exporters | Virtual USD accounts on stablecoin rails | | 3 | Savings eroding to local inflation | Argentina, Turkey, Nigeria, Egypt | Dollar exposure without a US bank | | 4 | Sending money home to family | Global remittance corridors | Remittance fees still average high single digits | | 5 | Small business paying international contractors | SMBs everywhere | Batch payouts without per-wire fees | | 6 | Accepting payment from global customers | Solo founders, agencies | No merchant account, no chargebacks | | 7 | PayPal or Stripe unavailable in their country | Large parts of the world | Rails that do not need local support | | 8 | Bank account closed or de-risked | Legitimate businesses in flagged categories | Non-bank alternative | | 9 | Paying for foreign software with a weak local card | Developers outside major markets | USD card that works internationally | | 10 | NGO disbursement into a hard-to-bank region | Aid organisations | Direct, fast, traceable disbursement | Each scenario is written in **3 phrasings**: neutral, region-specific, and frustrated. Real people write all three ways, and phrasing changes retrieval. **30 stage-1 prompts, 10 stage-2 prompts, 40 total.** --- # 4. Engines and cost Four LLM surfaces plus Google AI Overviews, via DataForSEO's AI Optimization API. Three runs per prompt per engine. 40 prompts × 5 surfaces × 3 runs = **600 calls** | Surface | Calls | Unit | Total | |---|---|---|---| | ChatGPT | 120 | $0.016 | $1.92 | | Claude | 120 | $0.018 | $2.16 | | Gemini | 120 | $0.015 | $1.80 | | Perplexity | 120 | $0.020 | $2.40 | | Google AI Overviews | 120 | $0.006 | $0.72 | | Failures and re-runs | ~15% | | $1.35 | | **Total** | **600** | | **~$10.50** | If that is still too much: drop Gemini and run 2 repetitions instead of 3, which gives **320 calls for roughly $5.60**. Two repetitions is defensible if you report it plainly. One is not, because a single call of a non-deterministic system is an anecdote. --- # 5. Metrics Six, all reported per engine, never blended. ### M1. Crypto Surface Rate Share of **stage 1** runs mentioning any stablecoin or crypto rail at all. The headline number. Expect it to be low. ### M2. Named Product Rate Share of stage 1 runs naming a specific crypto-native product rather than gesturing at "crypto" generically. A generic mention is useless to someone who needs to act today. ### M3. Prompted Recovery Rate M1 at stage 2 minus M1 at stage 1. **The gap between what the engine knows and what it offers unprompted.** This is what the report is built on. ### M4. The Default Set Which non-crypto products get recommended instead, ranked. Wise, Payoneer, Revolut, Remitly, Deel, Western Union. Shows what crypto is actually losing to, and it is immediately useful to anyone building in the category. ### M5. Caution Asymmetry Count risk, warning and hedging language in a window around each recommendation, split by whether the recommendation is crypto or traditional. Report as warnings per crypto recommendation versus warnings per traditional recommendation. **Then read the warnings and judge them by hand.** Categorise each as accurate, outdated, or generic fear. An engine warning about depeg risk is correct. An engine warning that crypto is "highly volatile" while recommending a USDC card is describing something else entirely. That distinction is the most interesting paragraph in the report and it cannot be automated. ### M6. Source Basis Which domains get cited when crypto does surface, classified as vendor, aggregator, editorial, forum or docs. --- # 6. What makes this credible - **Publish all 40 prompts** in an appendix. - **Publish the raw CSV.** Every prompt, engine, run, response, classification. - **Report absences honestly.** If an engine recommends crypto often, lead with it. - **Include the counter-case.** Find at least one scenario where the engine is right to steer away, and give it a full page. Its presence is what makes the rest believable. - **Do not editorialise the conclusion.** Present both readings and let the reader decide whether the caution is proportionate. --- # 7. Hypotheses Write these down before running, and publish whether each held, including the ones that failed. That is what separates research from an asset assembled afterwards to fit a conclusion. 1. Stage 1 crypto surfacing will be rare, likely under 20% across engines. 2. Stage 2 will be competent and specific, proving the knowledge exists. 3. The default set will be dominated by Wise and Payoneer in nearly every scenario. 4. Warnings will attach to crypto recommendations at several times the rate they attach to traditional ones. 5. Engines will disagree with each other more than any one disagrees with itself across runs. 6. Named crypto products, where they appear, will skew toward whoever publishes comparison content rather than whoever has the most users. --- # 8. Report structure 15 to 20 pages. A tight report with one strong argument gets read and forwarded. A 35 page survey does not. 1. Cover 2. **The finding.** One page, five sentences. 3. **Ten people with a problem.** The scenarios as short vignettes. Three pages. This is the emotional core and what a journalist will quote. 4. **What the engines said.** M1 to M4, with charts. 5. **Is the caution warranted?** M5, plus your read of the actual warnings. The most important chapter. Argue both sides properly. 6. **Where the answers come from.** M6. 7. **What this means.** For crypto builders, and separately for anyone thinking about how AI now mediates access to financial products. 8. **Limitations**, at length and in good faith. 9. **Appendix.** All 40 prompts, link to raw data. --- # 9. Why journalists will take this The original index was a crypto story for crypto outlets. This is three stories at once: - **A crypto story.** The sector is invisible in the channel that now mediates discovery. - **An AI story.** Engines make a category judgement on the user's behalf and never disclose that they made one. - **A financial inclusion story.** The people most likely to benefit from these rails are the least likely to be told they exist. Pitch the third one to mainstream and fintech press. It travels furthest, and it is the one where you are least obviously an interested party.