# Language attention on Steam — September 7 UTC / September 6 Brazil Authorized continuation of Steam research direction 6: identify games with substantial review attention outside English and little English-marked response. No causal claim about localization, buyer nationality, game quality or influence. **Completed result:** 456 games in the balanced primary sample, two supplementary observations, and scored-language subsets already available for 200 primary games. One page request timed out after 458 games; collection stopped without retry. The final primary sample uses the first 38 random selections from each of the twelve original strata. No failed game was replaced. All sampling weights were updated to population size / 38. The planned fresh 32-game API profile pass did not run; `available_profiles.py` uses only summaries already returned. See `collection_amendment.json`. All network collection is inactive. ## Sampling The frame is the independently collected September 5 catalog: paid, non-explicit games dated 2019 through August 2026 with at least 100 filtered reviews. There are 13,536 eligible games. Fully delisted/unavailable apps may be absent. Current free status, descriptors and release dates are imperfect historical measures. The frozen sample has 480 games: 40 randomly selected within each of twelve release-period x review-count strata. Periods are 2019–2022, 2023–2025 and January–August 2026. Review bands are 100–499, 500–1,999, 2,000–9,999 and 10,000+. Each member has weight `population_n / sample_n`. Names, descriptions, language support and any observed language shares did not enter selection. The first 24 games are a balanced feasibility pilot; they remain in the frozen sample. The public storefront's review-widget JSON includes global and English-marked Steam-purchase summaries for many larger games. Four independent API checks agreed within a predeclared one-percent or three-review tolerance. Small differences can reflect collection/cache or filtering differences and are preserved. Smaller games generally require API summaries for all and English; larger games with unavailable review props receive the same fallback. Store pages are requested with English language and US country. Individual review rows and account IDs are discarded; no media binaries are downloaded. The sample frame uses September 5 counts; new summaries are collected September 7 UTC. Keep both, with timestamps and count differences. These are a collection window, not an atomic snapshot. Missing values remain missing. A language summary with zero reviews is a measured zero, not a missing observation. ## Questions and thresholds - Fraction of games with at most 10% or 25% English-marked reviews, or a non-English majority. The unit is the game, with sampling weights. - English support versus no listed English support, keeping the latter's small sample explicit. Support flags say nothing about translation quality or when localization was added. - Differences by release period, review band, current price, and overlapping tags. Small tag/domain samples are exploratory, with sample sizes retained. - Discovery screen: at least 500 non-English reviews, fewer than 100 English reviews, and at most 10% English share. This separates relative language concentration from globally well-known games with large English totals. - Within-game English versus non-English recommendation proportions when both sides contain at least 50 reviews. Reviewers in different languages are selected populations; a difference is not a cultural or localization effect. The primary proportions use stratified weighted estimates. Approximate 95% sampling intervals use a design-based linearized ratio variance with finite population correction within the original twelve strata. These intervals do not incorporate source errors, language mislabeling, coverage bias, or temporal drift. Weighted domains use all sampled observations for variance estimation, including zeros outside the domain. Do not present raw sample proportions as population estimates. Review-weighted English share is a separate, secondary statistic and can be dominated by a few very large games. Boundary estimates with no observed variation have no reported interval; the plug-in zero variance is not interpreted as certainty. "English" means the language label attached to a review. It does not identify the reviewer's nationality, actual text language, preferred game language, residence, cultural identity, or first language. Review shares are not sales, players, revenue, awareness outside Steam, or localization return. ## Language profiles A planned second diagnostic sample would select sixteen games from the <=10% English-share bin and eight each from 10–50% and >50%, or all if fewer. Fixed random seeds are saved. These profiles deliberately overrepresent low-English cases and must not be pooled as estimates of language-market size. Each profile obtains all-language and English summaries plus Simplified and Traditional Chinese, Russian, Japanese, Korean, Brazilian Portuguese, European and Latin American Spanish, German and French. The unqueried languages remain an explicit remainder; measured shares are never renormalized to 100%. Language-specific counts are acquired serially and may not reconcile exactly if the source changes or caches differ. All discrepancies are retained. The storefront also exposes some language-specific score summaries. These are a selected subset and never a complete language distribution. They can prove that a named language supplies a majority when its count exceeds half the global count; missing entries cannot establish zero or weak attention. A planned separate adaptive discovery check would take the twenty largest cases meeting the strict discovery threshold, plus up to eight additional low-English-share build-tag examples with 1,000+ total reviews and at most 500 English reviews. Its purpose is to identify a majority review language for the illustrated games. It queries languages in a fixed order and stops once one measured language alone exceeds half the all-language count. Missing languages remain unmeasured. These cases are explicitly selected examples, not an additional probability sample. That script was not run after the timeout; no target manifest or new requests were produced for it. The completed discovery annotations use existing data only. The original sample contains seventeen games without listed English support; none has a wholly missing language-support array. A pre-outcome weighting check against the full frame's known English-support flag is retained in `sampling_benchmark.json`. No sample was replaced because of that comparison. Primary documentation: - https://partner.steamgames.com/doc/store/getreviews - https://partner.steamgames.com/doc/store/localization/languages ## Collection boundaries and reproduction Scoped cap: 1,400 requests including four initial probes, minimum two seconds between request starts, no retries, stop on HTTP or network errors, response size cap 4 MB. The research-root paused bulk-catalog policy remains unchanged. All new calls use a single locked collector. A `STOP` file or inactive scoped policy prevents further requests. Policies are closed at completion. Scripts: 1. `prepare.py` freezes the frame and sample; do not rerun to replace outcomes. 2. `collect_store.py 24`, `pilot_check.py`, then `collect_store.py 480` collect the primary summaries. These require the scoped collection policy to be active and are not needed to reproduce saved results. 3. `analyze.py` joins observations and computes primary weighted estimates. 4. `available_profiles.py` reconstructs the already returned scored-language subsets and annotates discoveries without any network work. 5. `charts.py`, `case_chart.py`, and `validate.py` reproduce the inspected figures and verification. The planned `prepare_profiles.py`, `collect_profiles.py`, `analyze_profiles.py`, and `discovery_languages.py` remain as unexecuted phase provenance and are not needed to reproduce the completed study. Use `../.venv/bin/python`. Requests, sanitized aggregate response records, sample probabilities and specifications remain available. Original raw catalog data and earlier studies are not changed.