2nd/10_Wiki/Topics/Architecture/리텐션(Retention).md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-리텐션-retention	리텐션 (Retention)	10_Wiki/Topics	verified	self	Retention User Retention Player Retention D1/D7/D30	none	A	0.9	applied	game-design product kpi growth monetization		2026-05-10	pending	language framework SQL/Python BigQuery / dbt

리텐션 (Retention)

매 한 줄

"매 리텐션은 game/product의 진정한 PMF(product-market fit) 신호다". 매 acquisition은 marketing이 만들지만 retention은 product 자체의 함수이며 매 D1/D7/D30 retention curve 의 shape 가 LTV·CAC·viability 를 결정한다.

매 핵심

매 Retention 정의

• Classic (N-day): 매 cohort 의 day-N 재방문 비율. D1, D7, D30 이 standard.
• Rolling N-day: 매 day-N±k window 내 재방문 — mobile 에 적합.
• Bracket: 매 누적 active in [N1, N2] window — mature product 에 적합.

매 모바일 게임 benchmark (2026)

• Hyper-casual: D1 35-40%, D7 10-15%, D30 3-5%.
• Mid-core: D1 45-55%, D7 20-25%, D30 8-12%.
• Hardcore RPG: D1 50-60%, D7 30-40%, D30 15-20%.
• Top decile: 매 D1 ≥ 55%, D7 ≥ 30%, D30 ≥ 15%.

매 Retention drivers

• Onboarding: 매 first-session 5-min 의 magic moment 도달 비율.
• Daily habit: 매 daily login reward, daily quest, energy regen.
• Social: 매 guild, friend list, leaderboard.
• Progression: 매 visible mastery curve (level, gear, collection).
• Live ops: 매 limited-time events, seasonal pass, FOMO.

매 응용

1. 매 D1 < 30% → onboarding/first-session 문제. 매 fix 우선순위 1.
2. 매 D1 OK but D7 cliff → core loop 단조 → progression/variety 부족.
3. 매 D30 OK but stagnant → endgame/social content 부족.

💻 패턴

Cohort retention SQL

WITH cohort AS (
  SELECT user_id, DATE(MIN(event_time)) AS install_date
  FROM events WHERE event = 'install'
  GROUP BY user_id
),
activity AS (
  SELECT e.user_id, c.install_date,
         DATE_DIFF(DATE(e.event_time), c.install_date, DAY) AS day_n
  FROM events e
  JOIN cohort c USING (user_id)
  WHERE e.event = 'session_start'
)
SELECT install_date, day_n,
       COUNT(DISTINCT user_id) AS retained,
       COUNT(DISTINCT user_id) * 1.0
         / FIRST_VALUE(COUNT(DISTINCT user_id))
             OVER (PARTITION BY install_date ORDER BY day_n) AS retention_rate
FROM activity
WHERE day_n IN (0,1,3,7,14,30)
GROUP BY install_date, day_n;

Retention curve fit (power law)

import numpy as np
from scipy.optimize import curve_fit

def power_law(d, a, b):
    return a * (d + 1) ** -b

days = np.array([1,3,7,14,30,60])
ret  = np.array([0.45, 0.30, 0.22, 0.16, 0.10, 0.07])
(a, b), _ = curve_fit(power_law, days, ret)
# extrapolate D90, D365
print(power_law(90, a, b), power_law(365, a, b))

LTV from retention + ARPDAU

def ltv(retention_curve, arpdau, horizon_days=365):
    return sum(retention_curve(d) * arpdau for d in range(1, horizon_days+1))

Daily login streak service

async function claimDailyReward(userId: string) {
  const u = await users.get(userId);
  const today = todayUTC();
  if (u.lastLoginDate === today) throw new Error('already-claimed');
  const isContinuation = u.lastLoginDate === yesterdayUTC();
  const newStreak = isContinuation ? u.streak + 1 : 1;
  const reward = REWARD_TABLE[Math.min(newStreak - 1, REWARD_TABLE.length - 1)];
  await users.update(userId, { streak: newStreak, lastLoginDate: today });
  await inventory.grant(userId, reward);
  return { streak: newStreak, reward };
}

Churn risk model (lightweight)

# features: days_since_install, sessions_last_7d, spend_total, level, social_count
# label: churned (no session in next 14 days)
from sklearn.ensemble import GradientBoostingClassifier
clf = GradientBoostingClassifier(max_depth=3, n_estimators=200)
clf.fit(X_train, y_train)
# segment top-K churn-risk users → targeted re-engagement push

Re-engagement campaign trigger

// trigger if user inactive 3-7 days, segment by spend, send personalized offer
async function triggerReengagement() {
  const candidates = await db.query(`
    SELECT user_id, last_session, total_spend, favorite_mode
    FROM users
    WHERE last_session BETWEEN NOW() - INTERVAL 7 DAY AND NOW() - INTERVAL 3 DAY
  `);
  for (const c of candidates) {
    const offer = c.total_spend > 50 ? 'whale_comeback' : 'soft_nudge';
    await push.send(c.user_id, OFFER_TEMPLATES[offer](c));
  }
}

매 결정 기준

상황	Approach
D1 낮음	Onboarding/FTUE 재설계, magic moment time-to-fun ↓
D7 cliff	Daily habit loop (login reward, quest), variety injection
D30 정체	Endgame, guild/social, seasonal pass
Whale retention	Personalized live-ops, VIP CS, exclusive events
New genre	Power-law extrapolation 의 X — empirical curve 만 trust

기본값: 매 D1/D7/D30 cohort tracking + power-law LTV proxy + churn-risk segmentation.

🔗 Graph

• 부모: Game KPIs · Product Analytics
• 변형: D1 Retention · D7 Retention · Cohort Analysis
• 응용: 디아블로_2(Diablo_II) · 클래시_로얄(Clash_Royale) · 알비온_온라인(Albion_Online) · 가차(Gacha)
• Adjacent: LTV · CAC · 부분_유료화(Free-to-Play) · 하이브리드_수익화(Hybrid_Monetization)

🤖 LLM 활용

언제: 매 retention curve 진단, segment 별 churn 원인 hypothesis, live-ops content idea, A/B test design. 언제 X: 매 specific causation 결론 (correlation only without experiment), 매 individual user prediction (privacy/legal).

❌ 안티패턴

• Vanity DAU: 매 retention 무시한 acquisition push → CAC payback 실패.
• D1 만 보기: 매 D7/D30 무시 → long-term LTV 왜곡.
• Linear extrapolation: 매 retention 은 power-law — 매 linear projection 은 LTV 과대평가.
• Forced daily login: 매 punishment-based streak → resentment + uninstall.
• Pay-to-retain: 매 monetization 위 retention 의존 → churn rate 가속.

🧪 검증 / 중복

• Verified (GameAnalytics 2025 benchmarks, Sensor Tower 2025 mobile gaming report, "Lean Analytics" Croll/Yoskovitz).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — D1/D7/D30 cohort + LTV + churn patterns