How the AI Citation Index is measured

What is measured

A fixed set of buyer questions runs against five AI answer surfaces every week. Each month we publish a snapshot of the accumulated results as a numbered edition. Brands are scored on how often and how prominently they are cited. Nothing here measures traffic, revenue or product quality — only citation.

Models sampled

• ChatGPT with web search (gpt-4o-mini)
• Google Gemini with Search grounding (gemini-2.5-flash)
• Perplexity (sonar)
• Google AI Overview (google-aio)
• Google AI Mode (google-ai-mode)

Categories

• Exchanges — 9 brands, 50 prompts
• Wallets — 9 brands, 50 prompts

What the questions ask about

The prompt sets are not a flat list. Each question is written against a buyer intent, and the balance between those intents is a choice that shapes every figure the index produces — a set weighted toward fees measures a different market than one weighted toward regulation. Both sets are published here so that choice can be argued with.

Scoring

A brand's score for one response is ((totalMentions - rank + 1) / totalMentions) * 100. A response that does not mention the brand scores zero. A brand's category score is the mean across every response, zeros included.

How rank is assigned

Rank is recomputed by us from the scores, not taken from the underlying measurement API. That API reports two rank figures that are not in the same rank space, and publishing either would place some brands in positions their scores do not support. We sort by score and break ties alphabetically, so the same data always produces the same table.

Known limitations

• AI assistants are non-deterministic: the same prompt can name different brands on different days. Weekly sampling, published monthly, reduces but does not remove this.
• Responses are sampled from a fixed prompt set. A brand strong on questions we do not ask will score lower than its real standing.
• Results are English-language and not geographically segmented.
• The brand set is curated. Entrants are added between editions, so a brand's first appearance is not evidence of a sudden rise.
• Model versions change underneath us. A shift between editions may reflect a model update rather than a change in the market.

Data source

Measurement is carried out with PeekaBoo. Editions are published by Cryptonist and each is kept live at a permanent URL.