Segmented targeting

Segmented targeting is a systematic approach to choosing who, what, when and through which channel to show, based on data on behavior, value and risks. Goal: Maximize incremental effect (income/retention) while minimizing harm and expense.

1) Segmentation goals and levels

Objectives: activation, monetization, retention, re-activation, cross-sell, RG/compliance interventions.

Levels:

Demographics/geo/device (basic).
Behavior/RFM (recency, frequency, monetary).
Value (LTV/margin).
Propensity (deposit, purchase, churn).
Uplift/sensitivity (probability of increase from contact).
Granularity: brand × country × platform × channel × content category - coordinate with traffic volume and budget.

2) Data and characteristics

Events: visits, clicks, registrations, games/bets, deposits/purchases, campaign responses.
Context: calendar (holidays/matches/salaries), attraction channel, application version, prices/limits.
Identities: user/device/IDFA/phone/e-mail; identity bridges.
Features: RFM windows (7/30/90), variety of content, hour/day of the week, ARPPU/frequency, responses to fluff/e-mail.
Quality: idempotency of events, dedup/antiboot, TZ = UTC storage + local views.

3) Segmentation methods

3. 1 Rules (rule-based)

Pros: explainability, speed, fail-safe.
Cons: fragility, local optima.

Example: "beginners, without a second visit 48h, mobile iOS."

3. 2 Cohort-behavioral (RFM/patterns)

Freshness × frequency × income classes.
Add behavioral clusters: favorite categories/time slots.

3. 3 Clustering/embeddings

k-means/DBSCAN/Gaussian mixes on normalized behavioral metrics.
Matrix factorization → interest clusters.

3. 4 Propensity/score-based

Event probability models: deposit, follow-up visit, apse; thresholds - from the cost of errors.

3. 5 Uplift/sensitivity

Contact gain models (T-learner, Uplift trees/GBM).
Зоны: Persuadables / Sure things / Do-not-disturb / Lost causes.
Uplift targeting gives the largest increment for the same budgets.

4) Design of targeting policies

Decision table

Segment/Condition	Context	Action	Channel	Cooldown	Guardrails
RFM: R7=0, F1, M0; `uplift_dep ≥ 0. 06`	beginners	welcome-offer S	push+in-app	3D	ROMI≥0
`churn_risk ≥ 0. 8` & `value_q ≥ 0. 8`	VIP	offer L/call	CRM+call	7d	zhaloby≤Kh
`RG_risk ≥ τ`	any	pause + RG tip	in-app	1d	FPR≤1%

Hysteresis: Input/output thresholds are different to avoid "blinking."

Quotas/rate-limit: per user/channel/segment; conflict policies - priority "security → economy → UX."

5) Experiments and causal evaluation

A/B/multi-user tests: for segments/policies/creatives; plan MDE, duration, stratification.
Quasi-experiments: DiD/synthetic control during regional rollouts.
Score: incremental revenue/withholding, uplift @ k, Qini/AUUC, ROMI, guardrails (complaints, RG).
Budget-pacing: Allocating budget to segments by expected margin return.

6) Channels and orchestration

Channels: in-app, push, e-mail, SMS, calls, onsite modules, personalized showcases.
Orchestrator: guaranteed delivery, retrai/backoff, idempotency 'action _ id', DLQ, priorities, quiet clock window.
Content/creatives: template library, A/B for creatives, frequency mouthguards.

7) Metrics and monitoring

Processes: coverage segment, latency (Decision→Action), deliverability, contact frequency.
Effect: revenue/retention increment, ROMI, uplift @ k, Qini/AUUC, NNT (how many contacts per 1 result).
Risks: complaints/unsubscribes/spam flags, RG indicators, fairness (difference in effects/errors by group).
Drift: PSI/KL by segmentation key features, share of "new" users outside known clusters.

8) Link to value and risk

LTV weighting: Prioritize segments by expected value rather than bare conversion.

EV rule:

[
EV = p_{\text{uspekh	action }\cdot\text {Value} -\text {Cost} - p_{\text{vred}}\cdot\text{Harm}
]

Contact if EV ≥ 0 and guardrails are normal.
Responsibility: RG/compliance takes precedence; explainability of the decision is mandatory.

9) Passports of segments and campaigns (templates)

Segment passport

Code: 'SEG _ RFM _ R0-7 _ F1-2 _ M0 _ Q3'

Definition and recipe (windows 7/30/90, anti-bots/fraud filters)

Size/update/freshness; intersection with other segments

Risks and exclusions (RG/compliance), owner, version

Campaign Passport

Objective and KPI (Incremental Revenue/Retention, ROMI)

Segment/Channels/Creatives/Frequency Caps

Assessment design (A/B or quasi-experiment), duration, MDE

Guardrails: zhaloby≤Kh, RG- flagi≤U, latency≤Z

Rollback Plan and Incident Runibook

10) Pseudo-SQL/Recipes

RFM segmentation (sketch)

sql
WITH acts AS (
SELECT user_id,
MAX(ts)                  AS last_act,
COUNT() FILTER (WHERE ts > NOW()-INTERVAL '30 day') AS freq_30d,
SUM(amount) FILTER (WHERE ts > NOW()-INTERVAL '90 day') AS money_90d
FROM user_activity LEFT JOIN payments USING(user_id)
GROUP BY 1
),
rfm AS (
SELECT user_id,
DATE_PART('day', NOW() - last_act) AS recency_days,
freq_30d,
money_90d
FROM acts
)
SELECT,
CASE WHEN recency_days<=7 THEN 'R0-7'
WHEN recency_days<=30 THEN 'R8-30' ELSE 'R31+' END AS R_bucket,
CASE WHEN freq_30d>=10 THEN 'F10+'
WHEN freq_30d>=3 THEN 'F3-9' ELSE 'F0-2' END    AS F_bucket,
CASE WHEN money_90d>=200 THEN 'M200+'
WHEN money_90d>=50 THEN 'M50-199' ELSE 'M0-49' END AS M_bucket
FROM rfm;

Propensity to deposit (train slice, no leaks)

sql
-- label: deposit in next 7 days
WITH snap AS (SELECT DATE_TRUNC('day',:cut) AS cut),
feat AS (... your RFM/behavioral features with condition ts <= cut...),
label AS (
SELECT u. user_id,
CASE WHEN EXISTS (
SELECT 1 FROM payments p
WHERE p. user_id=u. user_id
AND p. ts > cut AND p. ts <= cut + INTERVAL '7 day'
) THEN 1 ELSE 0 END AS dep_next_7d
FROM users u CROSS JOIN snap
)
SELECT FROM feat JOIN label USING(user_id);

11) Fairness, privacy and ethics

PII minimization: tokenization of identifiers, RLS/CLS, masking.
Fairness: check for differences in effects/errors across sensitive groups; Exclude invalid characteristics.
Transparency: reasons for targeting (top-features/rules) - available to support; path of appeal.
Frequency caps and "quiet hours" - against fatigue and harm to the user.

12) Anti-patterns

Segments "for beauty" without measurable increment.
Score by correlation rather than increment (no A/B/DiD).
Many overlapping segments → conflicting actions and spam.
Threshold solutions without hysteresis → "blinking" of contacts.
Lack of guardrails (RG/complaints/frequency) and explainability.
Without an online orchestrator - contacts are late, the effect is lost.

13) Segmented Targeting Launch Checklist

Goals, KPIs and budget are defined; consistent granularity
PIT data schemas and feature recipes; anti-bots/fraud filters are on
Segments are described by passports; conflict-matrix and priorities set
Selected method (rules/clustering/propensity/uplift) and evaluation design
Decision tables, hysteresis, mouthguards and rate-limit configured
Orchestrator and channels with idempotency and DLQ; "quiet hours"
Monitoring: effect (uplift/ROMI), risks (complaints/RG), drift feature
Documentation: segment/policy versions, owners, runbooks

Total

Segmented targeting is not a set of "shortcuts" in CRM, but a managed system: quality data and features → meaningful segments (value/behavior/sensitivity) → policies with hysteresis and guardrails → causal assessment of increment → stable orchestrator and monitoring. Such a system increases ROMI and LTV, reducing complaints and risks - and makes every touch appropriate, timely and ethical.

Segmented targeting