Where AI gets its answers

Reddit is the most-cited source in the corpus

Kunzum's Crypto AI Visibility Index recorded every source the engines cited across 480 answers to cross-border money questions, collected in September 2026. That produced 3,987 citations across 849 distinct domains. reddit.com is first, at 293.

No vendor, publisher, bank or regulator is cited more often than a forum. The twenty most-cited domains are below.

Editorial press is a rounding error

Grouping those 849 domains by what kind of source they are gives the allocation argument in one table. Vendor documentation dominates. Forum and social content is a clear second. Press coverage, the thing most crypto marketing budgets are pointed at, accounts for 26 citations out of nearly four thousand.

What that means if you want to be cited

Your own documentation and the threads people write about you are the corpus these engines draw on. A press release is not in the room. That is not an argument against press, which does other jobs, but it is an argument against expecting press to put you in an AI answer.

Half of ChatGPT's answers never touched the web

Enabling web search on these APIs permits retrieval. It does not force it. The Crypto AI Visibility Index recorded, for every response, whether the engine actually ran a search.

This matters for anything built on citation counts. Roughly half of what ChatGPT said in this study came from training rather than retrieval, which means no amount of publishing would have changed those particular answers. Citation work moves the retrieved half.

What the published literature says about citations

Kunzum’s citation counts sit alongside a small body of published work, and it is worth saying where they agree. Liu, Zhang and Liang’s 2023 paper Evaluating Verifiability in Generative Search Engines found that generative search engines routinely produce statements their own cited sources do not support. That is why this study logs a source URL for every response rather than trusting an engine’s account of where an answer came from. Yang’s 2025 analysis of news source citing patterns in AI search systems reports the same distribution shape the Kunzum index found here: a very long tail of domains, with a small handful taking most of the citations.

The retrieval-versus-training split has a documentary basis too. As of September 2026, OpenAI’s crawler documentation describes OAI-SearchBot as the crawler used “to surface websites in search results in ChatGPT’s search features”, and lists GPTBot separately as the crawler collecting training data. Perplexity’s crawler documentation states that PerplexityBot “is not used to crawl content for AI foundation models”. A site can be visible to one and invisible to the other, which is the distinction the web_search_ran column measures from the outside. The architecture underneath all of it is the retrieval-augmented generation design published by Lewis and colleagues in 2020.

None of this changes a number on this page. It does mean the numbers can be argued with on the same terms as the work they sit next to, which is the point of publishing a marked-up dataset alongside them. The Kunzum corpus is at the index dataset page, and the method behind it is written up in how a protocol gets named in an AI answer.