{
 "question": "How does geographic context (the country a question is asked from) change the brands ChatGPT and Gemini recommend for the same English buyer question, is the change larger than ordinary run-to-run variation, and how far does it move with changes in the cited sources?",
 "population": "40 country-neutral national buyer questions from study 7; ChatGPT and Gemini consumer apps via DataForSEO LLM Scraper, location set to the United States, United Kingdom, Canada and Australia, English. Block I: 5 runs per question, engine and country on 28 September 2026 (1,600 requests). Block E: the same questions ending \"in <country>?\", location set to that country, 2 runs (640 requests). Pilot: the single 26 September 2026 run (version 1.0, 320 answers).",
 "inclusion": "Answers of at least 200 characters.",
 "exclusion": "Unfinished answers under 200 characters (listed in stats.json).",
 "calculations": [
  "Brands: candidate strings from each answer (bold text, headings, ChatGPT brand entities), classified as brand or not (study 7 labels, plus Claude Opus for strings not seen before), name variants clustered per question, and counted where the normalized name appears in the answer. Order = order of first mention.",
  "Pair measures between two answers: Jaccard similarity of brand sets; same first brand; top-three overlap; rank-biased overlap (p = 0.9); Jaccard similarity of cited registrable domains.",
  "Within-country = mean over the 10 pairs of runs in each country; between-country = mean over the 150 cross-country pairs; gap = within minus between, per question and engine.",
  "Country share of variation: PERMANOVA-style R-squared of Jaccard distances among the 20 answers per question and engine grouped by country, with a 499-permutation p-value; chance level with 4 groups of 5 is 3/19.",
  "Geographic sensitivity index (set): 1 minus the share of majority brands (named in at least 3 of 5 runs in a country) shared by all four countries.",
  "Mixed model on pair-level Jaccard: same-country x engine, question random intercept (pairs share answers, so it is a check, not the primary inference).",
  "Source pathway (observational): brand overlap regressed on source overlap with question fixed effects; the same-country advantage in brand overlap before and after holding source overlap fixed, bootstrapped over questions (300 resamples).",
  "Local sources: share of cited domains with the country's code (.uk, .ca, .au; .us or .gov for the US). Local brands: share of an answer's brands whose model-coded primary market is that country.",
  "Explicit vs implicit: overlap between location-only and country-named answers in the same country, against location-only repeats.",
  "Expected geographic dependence of each question (low, medium, high) coded before analysis by Claude Opus, checked by Claude Sonnet; trend tested by OLS of R-squared on the ordinal code with engine, clustered by question.",
  "Intervals: 95% bootstrap resampling the 40 questions (2,000 resamples, seed 20260928)."
 ],
 "limitations": [
  "Country-level location set by the data provider, not a local account, city or device; account history is not varied.",
  "Two engines, one language (English), four countries, 40 questions, all consideration-stage \"best\" questions.",
  "Five runs on one day per condition; the pilot adds one earlier date. Longer drift is not measured.",
  "Observational: the source pathway is an association, not a controlled intervention on the sources.",
  "Country-code domains are a proxy for local sources; brand markets and question dependence are model-coded.",
  "Mention is not endorsement: a brand counts when named, including in a caution."
 ],
 "update": "Quarterly; next collection adds a no-location condition and city-level locations if the data provider supports them.",
 "@context": "https://schema.org",
 "@type": "Dataset",
 "name": "Same question, four countries: do AI recommendations change?",
 "version": "1.1",
 "dateCreated": "2026-09-26",
 "creator": {
  "@type": "Organization",
  "name": "Underneath",
  "url": "https://underneath.agency"
 },
 "url": "https://underneath.agency/research/ai-recommendations-by-country-study",
 "temporalCoverage": "2026-09-26/2026-09-28",
 "isAccessibleForFree": true,
 "license": "https://creativecommons.org/licenses/by/4.0/",
 "code": "cite/pipeline/ (collection, validation and analysis scripts)",
 "variableMeasured": [
  "collected",
  "pilot_collected",
  "reanalyzed",
  "bootstrap",
  "answers",
  "questions",
  "coding",
  "brands_per_answer_mean",
  "cells",
  "by_engine",
  "r2_chance_level",
  "engine_difference_chatgpt_minus_gemini",
  "mixed_model_pairs",
  "country_pairs_jaccard",
  "sources",
  "localization",
  "local_sources_vs_local_brands",
  "explicit_vs_implicit",
  "by_dependence",
  "dependence_trend",
  "low_dependence_cells_significant_pct",
  "by_industry",
  "brands_by_country",
  "date_check",
  "headline_balanced",
  "pilot_v10_statistics_reference"
 ],
 "dateModified": "2026-09-28",
 "research_questions": {
  "RQ1": "Geographic sensitivity: how much do recommended brands change when the same country-neutral question is asked from the US, UK, Canada and Australia?",
  "RQ2": "Stability: is the change between countries larger than the change between repeated runs from one country?",
  "RQ3": "Engines: do ChatGPT and Gemini differ in how much the country matters?",
  "RQ4": "Mechanism: do brand changes between countries move with changes in the cited sources, and with the share of local sources?",
  "RQ5": "Where: is the country effect larger for questions where the answer should depend on the country?",
  "RQ6": "Membership or order: does the country change which brands appear, which comes first, or both?",
  "RQ7": "Explicit vs implicit: how much of the localization produced by naming the country is produced by location alone?"
 },
 "framework": "Visibility is treated as a function of query, engine, date, user state and generation noise; this study varies one part of user state (country, with language held constant as English) and measures presence, first place, rank-weighted overlap, cited sources and local-market share separately.",
 "coding": "All labels are model-coded (Claude Opus primary, Claude Sonnet second coder, via claude -p); none are human-coded. Agreement figures are in stats.json under coding."
}