Best crypto wallet for beginners: what four AI engines actually answered

A week earlier we ran the same experiment on exchanges. That question has a default answer: Coinbase was named first in 46 of 100 answers, and in 29 of the 55 that named no brand. We expected wallets to look similar. They do not look remotely similar, and the reason turns out to be more interesting than the ranking.

Nobody wins this question

We sent twenty prompts, written the way a first-time buyer types rather than the way a marketer does. Eleven of them name no brand at all. Here is which wallet each of those 44 answers recommended first.

Read the shape rather than the order. The leader holds 8 of 44. Second place holds 5. Six wallets appear exactly once. Five answers decline to name a wallet and answer the category question instead, explaining hot against cold or custodial against self-custody and leaving the choice to you.

That is not what a settled category looks like. For comparison, on the beginner exchange question the leader held 29 of 55 brand-neutral answers. The wallet market, as these engines currently read it, has no default.

The engines disagree, and not slightly

The disagreement is not noise around a shared answer. The engines have different leaders.

Two things in there are worth sitting with. ChatGPT never recommends Zengo first on any of the twenty prompts, while Perplexity and Gemini do it seven times between them on the brand-neutral set alone. And every engine sometimes refuses the product question entirely and answers the category instead: Gemini does it twice on the eleven brand-neutral prompts, the other three once each.

If you are a wallet brand, that is the practical consequence: your position is not one number. It is four numbers that can point in opposite directions, and a single-engine check will mislead you in whichever direction that engine happens to lean.

Mentions are flat. First picks are not.

Being named anywhere in the answer is a much easier bar, and almost everyone clears it. Across the same 44 brand-neutral answers:

The five incumbents sit within three mentions of each other. Zengo is mentioned less often than any of them and recommended first more often than all of them, and Exodus is named in 14 of the 44 answers without being recommended first even once. That gap is the whole story of the page: presence in the answer is close to commoditised in this category, and the contest is entirely over the recommendation slot.

It also means a dashboard that reports share of voice as “mentioned or not” will show these six brands as roughly tied, which is true and useless. The metric that separates them is whether the answer tells a beginner to go and use you.

Where the answers come from

We recorded every source each engine returned: 1,592 citation slots across 251 distinct domains. Counted per slot, so a page cited three times in one answer counts three times, because that is how the engines expose it.

Then we classified each domain by whether it belongs to the product being discussed. Here is the result, and it is the most surprising thing in the run.

ChatGPT answers this category from the wallets’ own websites. 78.3% of its cited sources are product-owned, drawn from just 25 distinct domains across twenty answers. Perplexity, on the same twenty prompts, sits at 18.9% across 132 domains. That is roughly a four-fold difference in behaviour on identical inputs.

The pattern gets stronger as the question gets closer to a purchase. ChatGPT’s owned-source share runs 74.4% on the brand-neutral prompts, 79.5% on comparisons and 90.5% on the three brand checks. On “can I trust Trust Wallet” every source it returned was a trustwallet.com page. On “is a Ledger still trustworthy”, seven of its eight source slots were Ledger-owned.

What a wallet brand should take from this

We have to be careful here, because this finding cuts against our own headline number and we would rather say so than bury it.

With that stated, three practical readings hold up.

• Your documentation is answer surface, not just support. On ChatGPT, help centres and product pages are what the engine reads back to a buyer asking whether you are safe. If your security page, your recovery explainer and your fee page are thin, evasive or buried, you are not losing a support ticket, you are losing the answer. Write them to be quoted.
• On every other engine the third-party layer still decides it. Perplexity, Gemini and Claude pull between 12% and 24% from owned domains. For those three the picture matches the corpus: you are being described by other people’s pages. Optimising only your own site wins you one engine out of four.
• The recommendation slot is the metric, not the mention. Six brands are mentioned at near-identical rates. If your reporting stops at “we were mentioned”, you cannot see the only variable that is actually moving.

The third-party layer is narrow, and it is not the press

For the 73.7% of citations that are not product-owned, the distribution is tight. The top 20 of 251 domains carry 50.8% of every citation in the run.

Tier-1 crypto press is almost absent. CoinDesk, CoinTelegraph, The Block and Decrypt together account for under 1% of the citations in this run. Two review sites, CryptoSlate and Coin Bureau, carry more between them than every tier-1 masthead plus Forbes plus Reddit combined. That is consistent with what our wider corpus shows, where a 21-outlet mid-tier basket was cited 6.34 times more often than the tier-1 names.

Reddit lands at 1.88% here against 2.19% across the 39,948-event corpus. Close enough to call it stable, and small enough that anyone selling you a Reddit-first AEO programme should be asked to show their numbers.

Three names nobody has heard of

Worth flagging because it is the clearest evidence in the run that this category is winnable.

• Digitap was recommended first by Gemini on the lowest-fees prompt, over Trust Wallet, Exodus and every exchange.
• Zelcore was recommended first by Gemini on the free multi-chain prompt.
• walllet.com, a wallet on a four-L domain, appeared 18 times as a cited source and was named by Gemini on four separate prompts.

None of the three is a household name and none appears in a single ChatGPT, Perplexity or Claude answer in this run. Whatever they are doing on the source layer is working on exactly one engine. That is a fragile position, but it is also proof that the first-recommendation slot in this category is not locked by brand size. In the exchange study it effectively was.

How this was measured

Everything above is a count over a raw per-cell log. Nothing is modelled or estimated.

The scoring approach behind our own audits, including the engine weights and the mention, citation and authority split, is on the methodology page. The 39,948-event corpus referenced above is described in how AI picks its crypto sources.

What this does not say

• It does not say which wallet is best or safest. It reports what four AI engines said on one day. Zengo leading the first-pick count says its recovery model is well represented on the pages these engines read. It says nothing about the security of the product. Do your own diligence before moving funds into any wallet.
• It is not a rate. One sample per cell. A second run on the same prompts would move individual cells, and we would expect the shape to hold rather than the exact counts.
• It is one day. Source sets rotate fast. On a prompt we tracked in a separate study, ChatGPT replaced seven of its eight cited sources inside four days with no action from anyone.
• It is the US market. The two engines that require a location were set to the United States. The other two inferred it.
• We are not neutral about the topic. citeOS is our product and it sells AI citation audits to crypto brands, wallets included. The numbers here are counts anyone can re-derive from the method above, and we have published the finding that contradicts our own most-quoted figure rather than leaving it out.