86,094 AI citations: what gets crypto brands cited

What we measured

Between June and October 2026 we probed five AI engines with buyer-style prompts about 78 crypto brands, spanning exchanges, wallets, payment gateways, cards, DeFi protocols and infrastructure. Every URL in every answer was recorded and attributed to the brand the question was about.

That produced 86,094 citation events across 26,839 pages and 7,130 source domains.

A citation event is one engine naming one URL in one answer. A page cited in answers about four different brands counts four times. We did that on purpose. A page that serves fourteen brands is pulling more weight than one that serves a single brand, and we wanted to know which pages pull the weight.

We are publishing it because most advice about getting cited by AI is a vendor describing their own client roster. Very little of it is counted. Almost none of it is checkable. Our full scoring method is published in the open.

How concentrated are AI citations, really?

When I first looked at this I read it as hopeless. The engines pulled from 7,130 different domains and nothing accounts for more than 4.5% of citations. There is no Wikipedia of crypto answers. Nothing dominates.

Then I looked at the distribution.

Ninety-nine domains, 2.3% of the 4,395 in the dataset, decide half of every answer. At the other end, 2,218 domains were cited exactly once across four months. The median source domain in crypto AI answers has a single citation to its name.

So nothing dominates, and a hundred sources still decide the outcome. Both of those are true, and only the second one gives you something to do on Monday. Chasing presence across 7,130 domains is not a plan. Working a list of 99 is.

Where do the citations actually land?

Not on your website.

92.22% of citations point to pages the brand does not control. Own-domain pages account for 7.78%.

This is the number that most on-page AEO advice inverts. Schema markup, an llms.txt file, FAQ blocks and a clean docs site are all competing for roughly a twelfth of the available citations. They are worth doing, and they are not where the volume is. You can execute perfect on-page work and still be invisible, because the engines are mostly reading about you somewhere else.

If 92.22% of your citations live on other people's pages, citation growth is a distribution problem, not a website problem. That is the part worth sitting with.

Which single pages earn the most citations?

Six of the seven are the same thing: a ranked list of named products. Not a directory listing, not a press release, not a docs page. A list, with the brand's name in it.

The seventh is the interesting one. ledger.com's own "best crypto wallets" guide earned 100 citations, and Ledger is one of the 78 brands we audited. A vendor publishing a ranked list about its own category placed eighth overall. So the pattern is not "third-party pages win", it is "ranked lists win", and a vendor can publish one.

Across the full corpus, listicles account for 33% of citations. Among the most-cited pages we hand-read, that rises to 48%. Quote the second on its own and you will overstate the effect badly, so we give both.

Is being in a list better than having a page about you?

No, and this is where our own data cuts against the easy version of the headline above.

Measured per brand, being the subject of a page earns 6.22 citations. Being ranked in somebody else's list earns 3.78. Being absent from the page entirely but mentioned nearby earns 3.39. Subject-of-page wins, and the gap survives its confidence interval.

So a dedicated page about your brand is worth more than a slot in a list, per page. The reason lists still dominate the leaderboard is volume. There are far more "best crypto wallets" roundups in the world than there are pages written about any single brand, and one list can serve fourteen brands at once. The individual list slot is worth less. There are vastly more of them available to you.

Do both, and do not let anyone tell you on-page structure alone will close the gap.

Does more coverage reliably mean more citations?

Yes, and it is the most stable relationship in the dataset.

Citations track pages at roughly 1.48 citations per page that mentions a brand, correlation 0.80, coefficient of variation 9.5%. The number of pages you appear on predicts your citation count better than any single property of any one page.

It is a boring finding and it is the one I would keep if I had to throw the rest away. Nothing about tone, structure, schema or format came close to coverage breadth.

Is YouTube worth chasing?

For coverage, yes. For volume, no.

YouTube is cited for all 78 brands. No other domain in the dataset achieves full coverage. It accounts for 4.5% of all citations, the largest single source domain we have.

But those citations spread across 1,156 distinct videos, and the most-cited video in the entire corpus earned 17. The top ten videos together carry 7.3% of YouTube's total. The distribution is almost perfectly flat.

Compare that with the editorial roundup at 176 citations from one URL. On YouTube there is no single video to aim at, so you cannot concentrate effort there the way you can on a list. Publish so you are present. Do not build the plan on it.

What we would do with this

1. Stop sequencing your own site first. It caps at roughly 8% of available citations.
2. Identify the ranked lists in your category that engines already cite, and get into them.
3. Build a target list of around 100 domains and work it, instead of chasing presence everywhere.
4. Treat breadth of coverage as the primary metric, because it is the one that predicts.
5. Publish on YouTube for coverage, not for volume.

What this study does not show

Four limitations worth stating before anyone else states them for us.

The concentration is a power law, and power laws are ordinary. The top 1% of source domains carry 35.2% of citations, the top 5% carry 63.3%. Link distributions routinely look like this, so the shape is not the finding. The finding is that the head is small enough to enumerate and almost nobody has enumerated it. 99 is a workable number. Most teams behave as though it were 7,130.

The coverage relationship is close to definitional. More pages mentioning you produces more citations, which is not a discovery. What makes 1.48 usable is its stability: a coefficient of variation of 9.5% across brands of very different sizes. Use the constant as a planning number, not the correlation as a proof.

Our prompts shape our sources. We probed with buyer-style questions, and buyer-style questions surface buyer-shaped pages. Some of the listicle result is designed into the method. We think that is the right bias, because those are the questions that precede a purchase, but a different prompt set would produce a different source mix and we are not claiming otherwise.

We have not cut the data by month. The engines changed materially across the window. If you ask whether these shares have moved, the answer so far is no: the original five-month run and the live corpus agree within a point on every share in the box at the top.

One more thing we cannot explain. A small number of pages appear across an implausible number of brands, including an affiliate product page on a domain with a backlink rank of 9, cited for 23 different brands. Either the engines cite low-quality pages at a rate worth studying, or there is a contamination path in how we assembled the corpus. We do not know which, and we would rather say so than quietly filter it out.

Corrections and provenance

Every figure on this page was recomputed from the raw event file on 7 September 2026.

An earlier version of this dataset was published in July with category shares computed on a 4,969-event subset while carrying the full 86,094 label. Those rows have been recomputed. Two prior corrections, both August 2026, retracted a set of structural-lift multipliers that did not survive confidence intervals, and reversed a page-treatment ranking that had been inverted by a normalisation error.

We publish corrections because a score you cannot audit is a score you cannot trust. Full method at citeos.io/methodology.

Cite this

Emergence Media, "86,094 AI Citations: What Gets Crypto Brands Cited", citeOS, September 2026. Dataset: 86,094 citation events, 78 crypto brands, 26,839 pages, 7,130 source domains, June to October 2026.

Raw counts available on request for journalists and researchers: info@emergencemedia.agency.

That 92.22 per cent is an average, and we have since published a run where it inverts: on one engine in one category, 78.0 per cent of cited sources belonged to the brand being discussed. The counter-example is in the beginner wallet probe.

Common questions

How many brands were in the study?

78 audited crypto brands across exchanges, wallets, payment gateways, cards, DeFi protocols and infrastructure, measured June to October 2026. The dataset holds 86,094 citation events across 26,839 pages and 7,130 source domains.

What counts as a citation event?

One engine naming one URL in one answer. A page cited in answers about four different brands counts four times, because a page that serves fourteen brands is pulling more weight than one that serves a single brand.

Which AI engines were measured?

Five: ChatGPT, Perplexity, Gemini, Claude and Google AI Mode.

What share of AI citations point to a brand's own website?

7.78%. The other 92.22% land on pages the brand does not control, which is why citation growth is a distribution problem rather than a website problem.

Do listicles really earn more AI citations?

Across the full corpus listicles account for 33% of citations. Among the most-cited pages we hand-read, that rises to 48%. The single most-cited page in the dataset is an editorial best-crypto-wallets roundup at 176 citations, against 17 for the most-cited YouTube video.

Is being ranked in a list better than having a page written about you?

No. Measured per brand, being the subject of a page earns 6.22 citations, being ranked in someone else's list earns 3.78, and being absent from the page earns 3.39. Lists still dominate the leaderboard because there are far more of them and one list can serve fourteen brands at once.

Is this generalisable outside crypto?

Directionally. The coverage relationship and the concentration shape are likely general. The specific source mix is not, because aggregators like CoinGecko have no equivalent in most industries. Every brand in the study drew on a different set of source domains. The shape held, the specifics did not.

Can I get the underlying data?

Yes, on request, for journalists and researchers. Related reading: where AI gets its crypto answers and how AI picks which crypto projects to cite.