I run Kunzum, a one-person studio doing AI search visibility for crypto companies. Before that I spent six years writing crypto content and running marketing for protocols.
This post is the argument for why the studio exists, plus the two research findings behind it. Both studies are published with the raw data, so you can disagree with me using my own numbers.
The problem, stated plainly
Your buyers stopped starting at Google.
They open ChatGPT, Perplexity, Gemini, or they get an AI Overview before they see a single blue link. They type a question, they get four names, and they pick from those four. Nothing about your product quality enters that decision. What enters it is whether a retrievable document exists that answers their question and mentions you.
Crypto is unusually bad at producing those documents, for reasons that are mostly technical:
Client-side rendering. A huge share of crypto frontends are React apps that ship an empty shell and paint the content after hydration. Googlebot handles that reasonably well now. Most AI crawlers do not execute JavaScript at all. If your value proposition only exists after hydration, it does not exist.
Documentation treated as a support cost. Docs are usually the most factual, most structured, most chunkable content a crypto team owns, and they are maintained by whoever had time.
Unsettled category names. Is your product a crypto neobank, a stablecoin card, a payments app, or a wallet? You do not get to decide. The engines decide from published language, and if your category has four competing names your share of each one is thin.
Marketing spend that produces nothing retrievable. Threads, spaces, KOL posts, Telegram. All of it is real distribution and none of it is a document a model can retrieve six months later.
The net effect is that in crypto, being good is close to uncorrelated with being recommended. That is a testable claim, so I tested it.
Research one: what AI recommends when someone has a money problem
Full report and data: Does AI ever tell you crypto is the answer?
Most AI visibility research asks branded or category questions. "Best perp DEX." "Top stablecoin." Those questions are already won by whoever publishes comparison content, and the answers tell you little.
I wanted the harder version. Crypto has spent a decade pitching itself as the fix for broken financial rails. So I wrote prompts as the problems, not the products:
- I freelance for US clients and lose money on every transfer
- My savings lose value every year because of inflation
- My bank closed my account with no explanation
- I run a business in a country where Stripe does not operate
Ten scenarios like that. Forty prompts total, across five engines, three runs each, collected through the DataForSEO AI Optimization API so every response came back with its source annotations attached. No product names in any prompt. The question is whether crypto comes up unprompted, the way a solution comes up when it is genuinely part of the answer.
Across 360 unprompted answers, crypto was named as a usable, signup-able option six times.
Wise was named 246 times. Then PayPal, Payoneer, Stripe.
The pitch and the reality have separated completely. The industry says "banking the unbanked." When the unbanked ask, they get told to open a Wise account.
This is not the models being hostile. It is the models being accurate about what has been published. Every one of those scenarios has a well-written comparison page behind it, produced by a fintech company that wanted to be found. Crypto wrote threads instead.
A related pilot result, from two API calls that cost about four cents, shows the mechanism in miniature. Asked for the best crypto neobank for daily stablecoin spending on the same day within the same minute:
- ChatGPT named Coinbase Card, Crypto.com, Nexo, Binance Card, KAST, Gnosis Pay
- Perplexity named Bleap, KAST, Revolut, Xapo Bank
Nine distinct companies, one overlap. And Perplexity's top pick was Bleap, whose own comparison articles appeared three times in the nineteen sources Perplexity cited to produce that answer. The company that publishes the comparison is the company that wins the comparison.
Research two: where AI actually gets its crypto answers
Full breakdown: where AI cites crypto
The same collection run captured every cited source, which gave a corpus of roughly 4,000 citations to classify. This is the part that changed how I work.
Reddit was the single most-cited domain, at 293 citations.
Editorial press accounted for 26 citations. Out of roughly four thousand.
Vendor documentation carried the largest share of all, at 3,409 citations.
Sit with that ordering for a second, because it inverts the standard crypto marketing plan almost exactly. The press release you paid a retainer for is statistically invisible. The forum thread where someone asked whether your thing works, and got a straight answer, is load-bearing. Your docs are not a cost center, they are your primary retrievable asset and the thing engines quote when they describe you.
If you ship crypto software, the practical reordering looks like this:
- Make your docs answer buyer questions, not just integration questions. "How do I verify a Merkle proof" is covered. "Is this cheaper than a wire transfer for a 3,000 dollar invoice" is not, and that is the question being asked.
- Server-render anything that states what you are. If the crawler needs your JS bundle to learn your product category, you have lost before the ranking question even comes up.
- Make each page answer one question completely. Retrieval works on chunks. A chunk that reads well in isolation gets used. A chunk that depends on the three sections above it does not.
- Write the identifier into your own statistics. A stat sentence that gets retrieved without your name in it is a citation you gave away. I made this mistake on my own report page and had to fix it.
- Go where the citations are. Reddit beat every vendor in my data. I published a study proving that, then posted the findings to Reddit, because it would have been ridiculous not to.
There is also a caution asymmetry in the data worth reading in full. The engines hedge considerably harder on crypto than on the fintech alternatives for identical scenarios, and I hand-classified every warning with a verdict so you can see where the hedging is fair and where it is stale. That analysis lives at AI crypto warnings, and the list of what gets recommended instead is at what AI recommends instead.
The data is open
Every prompt, every raw response, the warning corpus with verdicts, and the scoring scripts are published at kunzum.xyz/research/data, with no email gate. Free to use with attribution.
I published it ungated for two reasons. The obvious one is that a report claiming "rescore it yourself" while hiding the CSV behind a form is not making that claim honestly. The less obvious one is that a gated file cannot be crawled, cited, or linked, which defeats the entire point of publishing data.
If you rerun it and get different numbers, I want to know. API responses are not identical to consumer app responses, answers drift week to week, and the limitations section says so at length.
What I am doing with this
Kunzum is AI search visibility for crypto, and only crypto. One operator. Technical SEO so crawlers can read you, a content pipeline aimed at the questions your buyers actually type, and mentions in the sources the engines pull from. The method is on kunzum.xyz with the client work that produced it.
Two other things I have built, both of which feed this:
- kymo tracks which AI crawlers hit your site, what they read, and which citations and clicks come back. Cookieless. I built it because I needed the measurement and nothing existed.
- RemoteStack is a remote job board, and mostly a live testbed for everything above.
Edition two of the index is running now, going one level deeper into crypto neobanks. Disclosure that belongs in the report and belongs here too: I am a user of one of the products in that category, and I will say so in the writeup regardless of how it scores.
I write about building these from Manali, at 2,050 metres, at narender.xyz. If you ship crypto software and want to know what the engines say about you, the fastest version is to just ask one and read the sources it cites. That takes ten minutes and usually ruins your afternoon.
Top comments (0)