# Crypto Card Geography Index (edition two)

Does an AI engine tell someone how to spend crypto as local money, and does the
answer change with the asker's country?

10 prompts x 9 markets x 3 runs across ChatGPT, Claude, Perplexity and Google AI
Overviews. 1,020 calls, $9.08, collected 9 September 2026. Zero unusable
records.

## Design

Prompt text is byte-identical in every market. Only the location changes:
`web_search_country_iso_code` on the LLM endpoints, `location_code` on the SERP
endpoint. Geography is therefore the single variable and any difference in the
answer is attributable to it.

Markets are the top crypto-adoption markets by the Chainalysis 2025 index,
excluding the US and UK as mature, with the US retained purely as a control:
India (1), Vietnam (4), Brazil (5), Nigeria (6), Indonesia (7), Philippines (9),
Turkiye (14), Korea (15).

Five scenarios, each asked twice: once with no crypto vocabulary at all, once
with stablecoins named. Stage one is what makes this research rather than a
leaderboard, because it measures what the engine volunteers rather than what it
can produce on request.

## Verified constraints

Claude rejects `web_search_country_iso_code` for Nigeria and Vietnam with a hard
`Invalid Field` error, so it has no coverage in those two markets. That is
recorded as absent, never as zero. ChatGPT, Perplexity and Google AI Overviews
cover all nine.

`max_output_tokens` is an invalid field on the ChatGPT endpoint. Sending it
produces a `40501` that looks like a country rejection and is not.

## Scoring corrections

Applied in `score.py`, never in `raw.jsonl`:

- Text is normalised before scoring. Markdown citation links are reduced to
  anchor text and bare URLs dropped, because ChatGPT embeds citation URLs inline
  and vendor domains would otherwise count as product mentions.
- Aliases are merged. "KAST" and "Kast Card" are one product; Wirex and "Wirex
  Card" co-occurred in all 17 of the latter's responses. Counting both inflated
  every card total.
- "Regulatory Grey Area" was matching Grey, the Nigerian fintech. Excluded by
  lookahead.
- Names that are also ordinary words are matched case-sensitively.
- Chains and tokens are tracked separately from products, since neither is
  something a person signs up for.
- Products are discovered from the corpus rather than from a watchlist. Reading
  the first pass surfaced Tria, Kolo, Pulsar, Bitnob, Cardtonic, Sphere,
  Jupiter Global, Bit2Me and Yellow Card, none of which were on any list.
- "Local" is defined by the data: any product with 70% or more of its mentions
  in a single market. Nothing was hand-picked as regional.

## Files

    prompts.json   the 10 prompt texts
    markets.json   9 markets, location codes, Claude coverage
    runner.py      collection. resumable, cost-capped, retries transient errors
    score.py       metrics
    raw.jsonl      1,020 records. Never edit by hand
    raw_full/      every unparsed API response, so a rescore costs nothing
    discovery.txt  open-vocabulary candidate list from the first pass
    results/       metric tables and summary