Duo Statistics: The Jynx Duo Index BR (Edition 2)

Between August 3 and October 14, 2026, Jynx recorded 482,181 duo-search likes from Brazilian players. The peak is at night: 22:00, 23:00 and 21:00 are the three busiest hours (22:00 leads, with 32,342) — and the 22:00 peak repeats game by game, from League of Legends to Free Fire. The mission is social: 91.75% of new players come looking for friends or mates (onboarding survey, n = 1,248), and 78.06% of players with games on their profile play at least one mobile title.

You may cite these numbers as "Jynx, Duo Index BR (aggregated data)" with a link to this page. Each study's methodology and limits are in the sections below.

Every number on this page comes from server-side aggregates computed from real Jynx app usage. Google Analytics is excluded from all analysis: it is pair-blind (UDA-004) and carries roughly 50% consent bias, so each metric is derived from server-side reads of the product's own collections.

Per the publication posture ratified in decision D29 (October 15, 2026; legitimate interest, LGPD Art. 7 IX): every number is aggregated and anonymized, each published cell has at least 50 observations, no cell combines 3 or more attributes, free text is never read, and every number is anchored to an explicit date window. Sub-floor cells are dropped, never padded. The Privacy Policy states the practice in its Data Usage section.

What is measured: like actions in the match queue by profiles attributed to Brazil — active duo-search behavior. The unit is the like because skips are never written: the histogram measures who acts, not who merely browses.

Window and volume: August 3 to October 14, 2026 (n = 482,181 likes). History starts August 3 because that is when the Firestore→BigQuery mirror went live; the one-off IMPORT backfill is excluded.

Timezone and stability: each like is bucketed in America/Sao_Paulo local time (UTC-3), keyed on the document's FIRST recorded created_at — the declined-like revival path re-stamps that field in place, and the MIN keeps runs deterministic.

Brazil attribution (with over-inclusion arms): the predicate is the house brShare convention — country OR UTC-03:00 timezone OR Portuguese language. The country field is null corpus-wide, so the timezone and language arms carry the selection, which over-includes non-BR UTC-03 and Portuguese-speaking users. The current-state bias (d0d1, issue 8) also applies: profile attributes are read as they are today, not as they were at like time.

Anonymization: hour-of-day cells, a single attribute; all clear the N ≥ 50 floor by orders of magnitude. Server-side analysis, no GA4.

Proxy definition (read before citing): a duo formed at matched_at "survives to day D" if the pair's private conversation last had a message at matched_at + D days or later. This is a conversation-continuation proxy — a LOWER bound on relationship survival: mates need not chat to stay mates, so the number reads as "duos still interacting", never as "duos still together".

Window and volume: March 4 to October 14, 2026. Corpus read: 54,390 relationships; analyzed: 53,445 (n); 26,460 without continuation evidence (no chat document or no last message); 0 block-censored observations.

Censoring: pairs blocked before day D leave the denominator — their D-day continuation is unknowable — unless they already messaged past day D (evidence beats censoring).

Evidence read: the last message is read as it stands at generation: a pair whose only post-D message lands after the window end still counts as survived — last_message_at is the LATEST message, not a within-window tally.

Edition-1 cut: cohort (UTC month of formation) × bucket (7d/30d/90d), nothing more — no rank cut (rank-at-match uninstrumented) and no game cut (no game-at-formation field). matched_at is authoritative with created_at as fallback; relationships with neither are excluded and counted in the notes.

Dropped cells (N ≥ 50 floor): the March–June 2026 cohorts fell below the denominator floor; the June (7d/30d/90d), July (30d/90d), and August 90d percentages were withheld for sub-floor survived subgroups. October's 7d figure (1.33%, n = 5,658) is a young cohort — read it with the evidence note above.

Pooled totals (edition 2): the total__7d/30d cells pool EVERY cohort old enough for the horizon (including individually sub-floor cohorts — the pooled denominator is the honest aggregate): day 7 — 5.84% of 44,645 duos; day 30 — 1.61% of 26,154. The pooled 90d has denominator 729 with the percentage withheld (surviving subgroup below the floor). Recent cohorts enter only the shorter horizons (right-censoring by cohort age).

What is measured: game_sessions documents created in the window — one row per session; origin is the enum recorded at creation (manual, auto-match, or community).

Window and volume: August 3 to October 14, 2026 (n = 1,181 sessions). Window totals: 954 manual and 227 auto-match.

Thin corpus, honest cells: monthly cells are UTC calendar months of creation; at 8–29 sessions a day, several monthly cells fall below the N ≥ 50 floor and are dropped, never padded — October's auto-match mix, for example, is not published. The whole-window totals (single-attribute cells) ride alongside the monthly trend on purpose.

Mirror: history starts August 3, 2026 (the Firestore→BigQuery mirror's live date); the IMPORT backfill is excluded.

What is measured: the onboarding intent question (users.user_intent), consented at collection and written once at onboarding completion — users who churn mid-onboarding have no intent recorded.

Window and volume: July 1 to October 14, 2026 (n = 1,221 unique respondents).

Multi-select: the question accepts several answers; the per-value count counts a respondent once PER selected value, and the unique-respondent total is published as a separate denominator — percentages are never silently renormalized.

No free text, closed enum: the free-text field behind the “other” option is never read. In this edition, 9 off-enum values were excluded (a client-writable field — values withheld as counts only). The social-mission cut (“find mates” + “meet people” and their union) publishes in this edition in its own section below; small n differences between the two studies derive from the latest-state mirror at each read moment.

What is measured: the same onboarding intent question, cut for the social mission: “find mates” (playing with people, duos and squads), “meet people” (making friends), and the UNION of the two — the share of new players who come looking for friends in either sense.

Window and volume: July 1 to October 14, 2026 (n = 1,248 unique respondents). Union: 1,145 (91.75%); find mates: 1,006 (80.61%); meet people: 920 (73.72%).

Multi-select: a respondent selecting both options counts once in each family and ONCE in the union — the union never exceeds the respondent denominator.

Anonymization: single-attribute cells with an explicit denominator sibling of at least 50 (the R1 floor); the “other” free-text field is never read; server-side analysis, no GA4.

What is measured: for every profile with at least 1 catalog game (n = 17,946), whether the player plays any mobile game (catalog platforms including Mobile/iOS/Android), any PC game (PC/Mac/macOS/Linux), both, or only one side.

Deterministic classification: each game's class comes from Jynx's own catalog (a list committed with the study: 179 mobile games, 612 PC games — cross-platform titles count as both; 27 console-only titles sit outside the classification; 53 profiles with only unclassifiable games were excluded and counted).

Overlap is the design: mobile (78.06%) and PC (88.30%) deliberately overlap — 66.66% play on both platforms; the disjoint cuts are mobile-only (11.40%) and PC-only (21.65%).

Client-writable field: game_ids is client-writable; off-catalog slugs (including manual games) classify no one — counted, never published. Latest profile state since March 4, 2026.

What is measured: the same unit as the timing study (match-queue likes by profiles attributed to Brazil), cut by the games of the player SENDING the like (profiles.game_ids). A like counts once per each game on the profile — multi-game attribution: per-game histograms share likes across a player's games and do NOT sum to a corpus total; each game carries its own total cell as its denominator.

Catalog allowlist: only published-catalog games reach cells (673 games with a total of at least 50 in this window); off-catalog slugs (a client-writable field) and manual (manual_*) games are counted, never named.

The parent study's caveats apply verbatim: likes-only (skips are never written), the 3-arm BR attribution (country OR UTC-03:00 timezone OR Portuguese, with over-inclusion), current-state attribution, and hour keyed on each document's first recorded created_at. São Paulo local time (UTC-3).

Window and volume: August 3 to October 14, 2026 (n = 4,832,893 game-attributed likes). Each game's duo page cites its own game's cells from this artifact.

What is measured: the shape of LFG sessions (game_sessions): the FINAL size (member_ids in the latest state — changelog rows re-log member joins, so the latest row is the final one) and the INTENDED size (max_members set at creation).

Window and volume: August 3 to October 14, 2026; 465 sessions analyzed (716 deleted before a final state and 221 with no members in the final state were excluded and counted).

Thin corpus, honest cells: sizes with fewer than 50 sessions drop — final sizes 3, 4, 5, 7 and 9; intended sizes 3, 6, 8 and 10 — never padded. Published: final 1 (154 sessions) and 2 (58); intended 2 (74), 4 (66) and 5 (226).

Right-censoring and honest reading: sessions still open can gain members — recent final sizes under-count. And “final 1” includes sessions that emptied after playing (members leave at the end): it is a final shape, not an abandonment rate.