Every answer, from all nine countries.

responses.csv

1,020 rows · 18 columns · 2.7 MB

One row per answer. This is the primary file: every reply the four engines gave, in full, with the prompt that produced it and the country it was asked from. A row is uniquely identified by prompt_id, market, engine and run.

Columns. prompt_id, scenario (1 to 5), stage (1 no crypto vocabulary, 2 stablecoins named), topic, prompt (the text sent), market (ISO code), market_name, adoption_rank, engine, model, run (1 to 3), timestamp, web_search_ran, mentioned_crypto, crypto_cards_named, all_products_named, n_sources, response_text (the complete answer).

roster.csv

227 rows · 7 columns · 8 KB

Every product named anywhere in the corpus, with how many answers named it, how that splits across the two stages, and how many of the nine countries it appeared in. Scoring was open-vocabulary: nothing was scored against a shortlist, which is how names like Kripicard and Kolo ended up in the study.

Columns. product, kind (card, venue, tradfi, local, rail, infra), responses, stage1, stage2, markets, market_list.

surface_rates.csv

68 rows · 8 columns · 2 KB

How often each engine raised crypto at all, and how often it named a card, by country and by stage. The headline rate on the report is computed from this table.

Columns. market, engine, stage, responses, mentioned_crypto, crypto_rate, named_a_card, card_rate.

localisation.csv

34 rows · 7 columns · 1 KB

Whether each engine answered as though it knew where the question came from: how often it named a product specific to that country, and how often it cited a domain from it. This is the file behind the finding that Claude never does either.

Columns. market, engine, responses, named_local_product, local_rate, cited_local_domain, local_source_rate.

share_of_voice.csv

732 rows · 6 columns · 20 KB

Who owns the answer in each country. One row per product per country, with the share of that country's answers naming it, and a flag for the products the localisation rule assigned to a single country.

Columns. market, product, kind, responses, share_of_market, market_specific_to.

sources.csv

2,668 rows · 5 columns · 94 KB

Every domain cited, per country and engine, with a flag for local top-level domains. Use it to see which publications an engine is reading when it answers someone in Lagos or Jakarta.

Columns. market, engine, domain, citations, is_local_tld.

prompts.json

10 entries · 2 KB

All ten prompts exactly as they were sent, so the study can be re-run against any engine. Five describe an ordinary money problem and never mention crypto. Five describe the same five problems with stablecoins named. No prompt names a country: the location was set through the API instead, which is what makes geography the only variable.

Keys. id, scenario, stage, topic, text.

markets.json

9 entries · 1 KB

The nine countries, their Chainalysis 2025 adoption rank, the Google location code used for AI Overviews, and whether Claude will answer as though it were there. Claude refuses Nigeria and Vietnam, which is why its figures rest on seven countries rather than nine.

Keys. name, aio_location_code, adoption_rank, claude, control.

runner.py

261 lines · 11 KB

The collection script. It is resumable, cost-capped, and skips Claude in the two countries it will not answer from rather than recording an empty reply as a null result. Every raw API response it received is in raw.jsonl unedited.

score.py

377 lines · 16 KB

The scoring script that produces every table above from the raw corpus. Its comments record each correction made during scoring, including the localisation threshold that first produced a false null for Brazil and Vietnam and how it was fixed.

raw.jsonl

1,034 lines · 4.3 MB

The unedited API responses, one JSON object per call, kept exactly as received. Nothing in this file was corrected by hand; all corrections happen in score.py, so this is what you re-score against if you disagree with any of it.

NOTES.md

67 lines · 3 KB

The study design and the running log: the two-stage structure, the market selection, the cost, and the two analyses that were discarded before publication because they turned out to be artefacts of our own counting.