Only 15% of AI Citations in Crypto Buying Answers Go to a Vendor's Own Site
By Zion Labs · Published August 4, 2026 · 9 min read
We ran two different measurements on the same 45 crypto and fintech companies. They disagree, and the disagreement is the finding.
Ask how often AI cites a brand’s own site in answers that already mention that brand, and the answer is 31.1%, stable across companies of wildly different size and technical quality.
Ask what AI cites when a buyer asks which product to choose, and only 15.4% of citations reach a vendor’s site at all. The rest go to Reddit, YouTube, review sites and aggregators.
Those are not the same question, and almost everything sold as AI visibility optimisation conflates them.
What we measured
Our Retrieval Readiness Score is a supply side measure: can AI crawlers reach a site, do its pages survive chunking, does it answer directly, does the brand resolve to a known entity. It says nothing about whether a site actually gets cited.
So we tested it twice, against two different definitions of getting cited.
Brand anchored. For 43 companies with coverage in an index of ChatGPT answers, we counted how many answers mention the brand and how many cite its own domain.
Category anchored. We wrote 35 buying questions, five for each of seven segments, and put each to ChatGPT, Perplexity and Gemini with web search live. 99 of 105 calls returned a searched answer. We recorded every domain each engine cited.
Finding one: the citation rate is close to a constant
Across 59,160 answers mentioning these brands, 18,382 cited the brand’s own site. That is 31.1% overall, and the median company sits at 31.8%.
A brand scoring 90 on readiness converts mentions into citations at roughly the same rate as one scoring 24.
Mentions and citations correlate with each other at 0.959. They are very nearly the same variable, which is why our first attempt to predict citation rate from readiness returned rho = −0.068. A dependent variable with almost no variance cannot be predicted by anything. That result was a property of the metric, not a fact about the industry, and we nearly published it as one.
Finding two: readiness moves mentions, once you are already in the conversation
Measured against mentions and citation volume rather than the near constant rate, and controlling for brand size, readiness does show up:
| Relationship | Raw | Controlling for organic traffic |
|---|---|---|
| Readiness to AI mentions | +0.368 | +0.476 |
| Readiness to citation volume | +0.318 | +0.436 |
| Readiness to citation rate | −0.068 | −0.075 |
| Organic traffic to AI mentions | +0.831 |
Both readiness figures get stronger once brand size is held constant, which means the raw numbers were understating the relationship rather than inflating it.
Brand size is still the larger factor. But readiness carries a real independent effect, and it is the part a team can change this quarter.
Finding three: none of that gets you into a buying answer
Then we asked the questions a customer actually asks. Readiness against category citations: rho = 0.117. Nothing.
25 of the 45 scored companies were cited zero times for questions in their own segment. That includes some of the highest scorers in our index, at 89, 86, 86, 85 and 84.
Here is where the 661 citations went instead:
| Domain | Citations | Share |
|---|---|---|
| reddit.com | 34 | 5.1% |
| youtube.com | 24 | 3.6% |
| nerdwallet.com | 13 | 2.0% |
| coingecko.com | 12 | 1.8% |
| forbes.com | 9 | 1.4% |
Spread across 345 different domains, with only 15.4% reaching any vendor site.
When someone asks which wallet to buy, the engine is not evaluating wallet websites. It is reading what other people said about wallets. A perfectly structured product page does not enter that calculation, because the engine never went looking for one.
The companies we could not score are doing fine
Ten companies were excluded from our readiness index because their sites turned our crawler away. We have always insisted that unscored means we could not see, not that a company failed. This run tested that.
Koinly was cited in 4 of 5 questions in its segment. Coinbase in 3 of 5. Both are among the strongest performers in the entire sample, and neither could be measured.
A crawler being blocked tells you about the crawler. Any index that scores those companies as zero is publishing an artifact of its own tooling.
What this means in practice
The chain that holds up is narrower than the industry’s pitch:
Retrieval readiness increases how often AI mentions you, and citations follow mentions at a fixed rate of about one in three. It does not put you into answers where the engine was never looking at vendor sites to begin with.
Readiness compounds visibility you already have. It does not create visibility.
If your brand is already discussed, the technical work pays. If you are trying to enter category buying answers, the lever is being present in the sources those answers are built from, which in this industry means Reddit, YouTube and the review and comparison sites. That is closer to digital PR and community presence than to schema markup.
We sell the technical work. It is not the thing that gets you into a buying answer, and we would rather say so.
Limitations, stated properly
Our buying questions lean toward comparison intent. Most begin with “best” or “cheapest”, and that is exactly the query class where Reddit and review sites dominate. A prompt set weighted toward brand qualified questions such as “is Trezor safe”, or toward problem and how to queries where documentation is the answer, would likely show vendor sites doing better. This finding is strongest for the discovery stage and should not be read as covering all AI search.
Both measurements are cross sectional. They compare different companies at one moment and cannot establish that improving readiness improves outcomes. Only measuring the same companies twice can do that. We intend to.
Small samples. 43 companies in the brand anchored analysis, 45 in the category one, 35 questions, one vertical. 25 of 45 companies scored zero category citations, which weakens that correlation for the same reason the near constant rate weakened the first one.
One index, three engines. Mention data is indexed ChatGPT answers, United States English. The category measurement covers three engines live, which is better, but a single run is a snapshot and engines change.
Correlation, not causation. Well built sites and well known brands are not independent. Controlling for organic traffic reduces that confound without eliminating it.
Method
Companies were drawn from our Crypto and Fintech Retrieval Readiness Index across seven segments: centralised exchanges, DeFi protocols, layer 1 and layer 2 chains, wallets, neobanks and fintech apps, payments and on ramps, and crypto tax tools. 55 measured, 45 scored, 10 unmeasurable.
Category questions were run on 4 August 2026 against ChatGPT, Perplexity and Gemini with web search enabled. Any response where the model did not actually search was discarded rather than counted as a non citation, since an answer written from memory says nothing about retrieval. Google’s vertexaisearch.cloud.google.com redirect wrapper was excluded from citation counts as a transport artifact rather than a source.
Companies were credited only for questions asked of their own segment. Correlations are Spearman rank, chosen because traffic and citation counts are heavily skewed. Partial correlations control for estimated monthly organic traffic.
We are publishing this without the company level readiness scores. The graded index is a separate publication held to a separate standard, because putting letter grades on named companies deserves more scrutiny than an aggregate does.
Next
Re running both measurements against the same companies in 90 days, which is the only route from correlation to causation. Widening the question set to cover brand qualified and problem intent rather than comparison intent alone. Extending beyond one vertical.
If the readiness effect holds longitudinally, it is a real lever and we will say so. If it does not, we would rather be the ones who found out.
Zion Labs researches how AI answer engines choose what to cite, and builds the tools and monitoring to help brands become the source they trust. Try the free tools or book a free audit.
