What AI tells a developer or a business choosing blockchain infrastructure
A developer choosing an RPC provider or a bank choosing a platform now asks an assistant for a shortlist. We measured what those answers are built from, because the shortlist is decided there before anyone visits your site.
On 23 October 2026 we put 12 questions to Perplexity and to ChatGPT with web search, the kind a developer or a business asks before choosing: the best RPC provider for Ethereum, which chain to build on, the best Layer 2 for a consumer app, which oracle a DeFi protocol should use, the best platform for tokenising real-world assets, the best data API, the best institutional custodian, which enterprise platform a bank should use, the best embedded wallet SDK, indexer and stablecoin payments API, and the most trusted audit firm. Every question asked for a recommendation, 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 | 72% on Perplexity (86 of 119 citations), 74% on ChatGPT (17 of 23) | The shortlist is built on comparisons. Every infrastructure category has a few “best X” pages that decide who gets named; being on them, described correctly, is a channel of its own. |
| Citations on the own site of a provider the answer recommends | 14% on Perplexity (17 of 119). In 12 of those 17 the page was the provider’s own “best of” list or comparison; 2 were its developer docs | The providers that get cited for their own recommendation are mostly cited for their own comparison page. A fair, well-made “best X” or “X vs Y” page on your domain is the page most worth building. |
| Coinbase, the product brand named most often | Named in 5 answers across both surfaces, for custody, its wallet SDK and the Base network. coinbase.com cited in none of them | Even the biggest company in the answers is recommended on other people’s pages. Size does not put your site into the answer; the pages that compare you do. |
| Citations on another company’s blog, learn or docs pages | 34% on Perplexity (41 of 119), in all 12 answers. 14 more citations were development agencies’ blog posts | Answers lean on other companies’ blogs, including development agencies writing lists of providers. Every category has an explainer seat that a provider can take on its own domain. |
| Developer docs and GitHub cited | 6 of 119 on Perplexity, none of them GitHub. None of 23 on ChatGPT | Docs are almost never what the answer quotes. They convert developers who already found you; comparison and explainer pages are what get you found. |
| Answers citing media: crypto, tech or business press | 7 of 12 on Perplexity, but 13 of the 14 media citations are the outlet’s own ranked list, not news. 1 of 12 on ChatGPT | Media count here through their ranked lists. A trade publication’s “top providers” list moves the answer more than a funding announcement does. |
| ChatGPT answers given without searching the web | 5 of 12: which chain, which Layer 2, which oracle, which enterprise platform for a bank, which indexer. All 7 that searched added “2024” to the query | On chains, Layer 2s, oracles, enterprise platforms and indexers, ChatGPT answered from memory. Only a lasting record in the sources it learned from changes what it says about your category. |
Asked which enterprise blockchain platform a bank should use, ChatGPT answered without searching and recommended Microsoft Azure Blockchain, a service Microsoft retired in September 2021. Asked for the best Layer 2 for a consumer app, it did not mention Base, which Perplexity put first. A model repeats what it learned, and in infrastructure what it learned is often two years old.
Method: Perplexity (sonar) and ChatGPT (gpt-4.1-mini with web search) through the API, 23 October 2026, one answer per question, US. 22 questions in two sets: 12 a developer or a business asks before choosing blockchain infrastructure and 10 a collector or player asks. No question names a brand. 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: 10 to 12 answers a set on each surface, and Perplexity returns at most ten sources an answer, so its ten is a ceiling, not a finding.