What AI tells a web3 user before they connect a wallet
Users of onchain products ask an assistant which protocol, wallet or bridge is safe. We measured which pages the answers are built from, because that, not your landing page, decides who gets the wallet.
On 23 October 2026 we put 36 questions to Perplexity and to ChatGPT with web search, the kind a user asks before choosing: the safest exchange for a beginner in the US, the UK or the EU, the most secure hardware wallet, where to earn yield on stablecoins, which lending protocol is safest, where to stake ETH, how to tell whether a new token is legit, the best crypto debit card, the best crypto tax software. Every question asked for a recommendation, so every answer names products, and we read which pages each answer was built from.
| What we measured | Result | What it means for you |
|---|
| Cited pages that are a ranked list or a comparison | 69% on Perplexity (246 of 358 citations), 73% on ChatGPT (56 of 77) | The decision is made on lists and comparisons. For a protocol or a chain, being on the lists that rank, and described correctly there, is a channel of its own. |
| Citations on the own site of a project the answer recommends | 6% on Perplexity (22 of 348). In 15 of those 22 the page was the project’s own “best of” list ranking itself among competitors | Your own site rarely backs the recommendation. Docs and landing pages are not what answers quote; comparison and explainer pages are, and most protocols do not have them. |
| Coinbase, the brand named most often | Named in 12 Perplexity and 16 ChatGPT answers. coinbase.com cited in none of them | Even the category leader is recommended on other people’s pages. Total value locked or a famous token does not put your site into the answer; the pages around you do. |
| Citations on another crypto company’s blog or learn pages | 30% on Perplexity (105 of 348), in 32 of 36 answers. A tax-software company is the fourth most-cited domain, for its pages on the best exchanges | Other companies’ learn pages are where answers are sourced: an exchange explains your protocol, a tax tool ranks your category. That seat is open to you on your own domain. |
| Answers citing crypto media | 27 of 36 on Perplexity, but 32 of the 50 media citations are the outlet’s own ranked lists, not news. 2 of 36 on ChatGPT | Crypto media count here through their ranked lists more than their news. Coverage that ends up in a list moves the answer; a launch announcement rarely does. |
| ChatGPT answers given without searching the web | 11 of 36, including both stablecoin-safety questions and all five questions about whether a new token is legit | On the riskiest questions, such as whether a new token is legit, the model answered from memory. Only a lasting, accurate record in the sources it learned from changes that. |
| Reddit, YouTube and regulators cited | Reddit 8 citations of 358, YouTube 0, regulators 2, on Perplexity. None of the three on ChatGPT | Threads and regulators barely appear in these answers. Threads still matter in Google and in a user’s own research, which is where we use them. |
Asked whether a new token was legit, ChatGPT did not search once in five questions. It answered from what it had learned. For a web3 product that means the record about you, on the lists, the dashboards and the coverage, has to exist before the user asks, because the model will not go looking for it.
Method: Perplexity (sonar) and ChatGPT (gpt-4.1-mini with web search) through the API, 23 October 2026, one answer per question, 36 questions in eight categories, US unless the question was about the UK or the EU. A source is a distinct page the answer cites; ChatGPT’s logo links above each product it lists are not counted. Domains were classified by hand and five answers checked against the raw responses. A small sample, read as a direction: Perplexity returns at most ten sources an answer, so its ten is a ceiling, not a finding.