Method
Four steps, run in the same order every time. Each one below carries the measured evidence it rests on, taken from Kunzum's own published research rather than from a framework.
Getting named in an AI answer is not one thing. It is a retrieval problem, a wording problem and a structure problem, and they have to be solved in that order, because fixing the wording of a page nobody retrieves changes nothing.
What follows is the whole method. There is no account layer and no team you never meet. The person who writes the pages is the person who reads the results.
Before anything is written, your buyers' real questions get run across four engines and the answers are written down. Who gets named, in what order, and which source each answer came from. That becomes the baseline, and it never changes, so every later month is measured against the same start rather than against a moving target.
The questions come from your side of the business, not from a keyword tool. What the BD lead answers twice a week. What people ask in the Telegram before they commit capital. Keyword volume is a poor proxy here, because nobody types a keyword into a chat window.
Every question is asked at least three times per engine. This is the part most audits skip, and it is why most audits are wrong.
In Kunzum's Crypto AI Visibility Index, three runs of the same prompt on the same engine disagreed with each other in 20% of cases. Across 40 prompts, 13 were total splits: one engine raised a topic in all three runs while another raised it in none. A single call to a non-deterministic system is an anecdote.
What buyers call your product is rarely what your site calls it. The site says infrastructure and the buyer says the thing that lets me get paid. Closing that gap is the single largest fix available on most crypto sites, and it costs nothing but attention.
The words come from Telegram, Discord and sales calls, not from a brainstorm. Then the pages get rewritten to answer in those words, in the same sentence as the claim, so a model matching a buyer's phrasing finds a page that already speaks it.
This sounds soft until you watch the exact string change the answer.
Asked for the world's first bitcoin perp dex, Google AI Overviews names Velar PerpDex. Asked with the exact string perpdex, with no spelling correction applied, the same engine returns a general history of perpetual exchanges and never mentions Velar at all. One space, two different answers, same question.
Models do not read your page top to bottom. Retrieval pulls one passage and judges it on its own, without the heading above it or the paragraph after it. Most crypto sites put the claim in one block and the proof three blocks down, which means the passage that gets retrieved is the one that cannot support itself.
So every block is rewritten to stand cold. The claim and its evidence go in the same place. Statistics carry their own attribution, because a number with no source attached is a number a model will not repeat. Pronouns get replaced with the actual subject at the start of each section, since the referent lives in a block the retriever never saw.
None of this makes the writing worse if it is done properly. It reads as confidence rather than repetition, because a sentence that names its own subject is a sentence that sounds sure of itself.
This site is written to the same rule. Every headline statistic in the research names the study and the date inside the sentence that carries it, so a passage lifted out of context still says where it came from.
Rewriting your own pages is necessary and not sufficient. Engines assemble answers from documentation, aggregators, and the places people discuss your category. So the work extends off-site: the docs, the comparison pages, the threads buyers actually read.
Which places, specifically, is not a matter of opinion. It was measured.
Across 3,987 citations in the Crypto AI Visibility Index, vendor documentation accounted for 3,409 and editorial press for 26. reddit.com alone was cited 293 times, more than any vendor, publisher, bank or regulator in the study.
The engines also read different internets. Perplexity drew on 447 distinct domains and cited Reddit 222 times on its own. Claude drew on 174. Optimising for one engine is not optimising for four.
Then the same questions run again every month, in kymo, a tracker built here because nothing on the market recorded which crawler read what, or which citation actually sent a human.
It does not put you in an answer the engine never retrieves for. In the index, ChatGPT ran a web search on 54% of its answers and Claude on 65%. The rest came from training data, and no amount of publishing would have changed those particular responses. This work moves the retrieved half.
It does not work as a one-off. The Velar check returns a different answer depending on a single space in the query, and the engines change what they say between runs on the same day. A position in an AI answer is held rather than won.
It is also not a substitute for having something worth citing. Every engine in the research was quoting a source. If nothing accurate and specific exists about your product anywhere, the first job is to write it, not to optimise it.
Keep reading