Build pillar · Loyalty-journey agent · Per-member per-location LTV math · Published August 29, 2026

How to build per-member per-location LTV math for multi- location loyalty

An implementation architecture for joining the operator loyalty, POS, ecommerce, CRM, lifecycle, subscription, reservation, and booking data into a per-member per-location LTV computation with the operator-data-science-team-approved math methods, cold-start, refresh cadence, cross-banner handling, anomaly detection, and NBA handoff — gated by the per-member compliance overlay covering Robinson- Patman, ECOA Reg B, Fair Housing Act, EU AI Act Article 22, and the wider regulatory surface.

Tier 1 — AI Readiness Assessment ($10k)Tier 3 — Fractional CMO

Or take the 3-question shape diagnostic first — no email required.

What you will build

A data-source catalog joining loyalty (Square Loyalty, Smile.io, LoyaltyLion, Yotpo Loyalty, Annex Cloud, Punchh, Paytronix, Marigold Loyalty, Tango Card, Antavo, Open Loyalty, Loopy Loyalty, Stamped Loyalty, Toast Loyalty, Lightspeed Loyalty, Como), POS (Square POS, Toast POS, Clover, Lightspeed, Aloha, Revel, TouchBistro, ShopKeep, Vend), ecommerce (Shopify, BigCommerce, WooCommerce, Magento, Squarespace Commerce, Wix), CRM (Salesforce, HubSpot, Microsoft Dynamics 365, Zoho, Pipedrive), lifecycle (Klaviyo, Iterable, Braze, Customer.io), subscription (Recharge, Bold Subscriptions, Stripe Billing), reservation (OpenTable, Resy, Tock, SevenRooms), and booking (Square Appointments, Vagaro, MindBody, Booker) data into the operator data warehouse (Snowflake, BigQuery, Databricks, Redshift, or Postgres).
An LTV class and math-method enginecovering naive historical, cohort-based average, BTYD (BG/NBD, Pareto/NBD, MBG/NBD), Gamma-Gamma monetary value, hierarchical Bayesian (PyMC, Stan, NumPyro, bambi, brms, RStan with arviz), survival analysis (Cox PH, Kaplan-Meier, Weibull AFT, Random Survival Forest, DeepSurv, DeepHit), gradient-boosted trees (XGBoost, LightGBM, CatBoost, NGBoost, EBM), neural networks (MLP, RNN, LSTM, Transformer), customer-equity Markov, per- location attribution, and per-attribution-window (30-1,825 days).
A cold-start, refresh-cadence, and cross-banner handling layer using first-purchase priors, look-alike imputation, Bayesian shrinkage to portfolio mean, a 30-day minimum observation window, daily-to- quarterly refresh logic plus event-driven triggers, per- banner identity stitching, per-portfolio member IDs, cross-banner earn-redemption, and FDD Item 12 territorial protection.
An anomaly detection and NBA handoff with per-member z-score (versus cohort and versus historical self), sudden-spike fraud signals, sudden-drop churn risk, per-cohort drift, and routing into the loyalty-journey decisioning siblings.
A per-member compliance gate as policy- as-code (OPA Rego, AWS Cedar, Casbin, Cerbos, or Oso) covering Robinson-Patman, ECOA Reg B disparate impact, Fair Housing Act, EU AI Act Article 22 plus GDPR Article 22, FTC fake-review rule and Negative Option ROSCA, HIPAA, FCRA, GLBA, CCPA/CPRA, GDPR, FDD Items 12/17/19, cannabis state board, alcohol DISCUS tied-house, and the rest of the operator regulatory surface as enumerated in the FAQ.
An audit trail and end-to-end replay to operator-controlled WORM storage at per-statute retention windows with deterministic replay.

How the architecture is built

Data-source ingestion. Inventory which loyalty, POS, ecommerce, CRM, lifecycle, subscription, reservation, and booking vendors the operator runs. Stitch identity through the operator CDP (Segment, mParticle, Rudderstack, ActionIQ, Treasure Data, Tealium, or Lytics) where one is wired. Stream into the operator data warehouse with per-source freshness tracking.
LTV class taxonomy. Author the per- member LTV class set with operator marketing, finance, data science, and counsel.
Math-method engine. Stand up BTYD via the lifetimes library or PyMC-Marketing where non-contractual fits; hierarchical Bayesian on PyMC, Stan, NumPyro, bambi, or brms where partial pooling helps small-sample locations; survival via lifelines, scikit-survival, DeepSurv, or DeepHit where time-to-churn is the target; GBT via XGBoost, LightGBM, CatBoost, NGBoost, or EBM where large feature sets call for it; neural networks where data volume justifies it; customer-equity Markov for tier-transition modeling.
Cold-start engine. Wire first-purchase priors, look-alike imputation, Bayesian shrinkage, and the 30-day observation window.
Refresh-cadence engine. Wire daily, weekly, monthly, quarterly, and event-driven refresh logic through the operator workflow engine (Temporal, Inngest, Trigger.dev, Vercel Queues).
Cross-banner handling. Stitch identity across banners with operator-counsel-approved FDD Item 12 territorial-protection respect.
Anomaly detection. Stand up per-member z-score, fraud-spike, churn-drop, and cohort-drift detectors.
Compliance gate. Author the policy-as- code library with operator counsel approval per framework.
NBA handoff and audit trail. Route to the loyalty-journey decisioning siblings. Persist every computation to operator-controlled WORM storage at per- statute retention windows.
End-to-end replay and cross-skill handoffs. Build deterministic replay; coordinate with the 40+ sibling skills enumerated in the FAQ.

Frequently asked questions

What does per-member per-location LTV math do that a loyalty-platform native LTV column does not?

Loyalty-platform vendors (Square Loyalty, Smile.io, LoyaltyLion, Yotpo Loyalty, Annex Cloud, Punchh, Paytronix, Marigold Loyalty, Tango Card, Antavo, Open Loyalty, Loopy Loyalty, Stamped Loyalty, Toast Loyalty, Lightspeed Loyalty, Como) ship strong points-balance and earn-rate primitives plus a per-store LTV figure that typically sums historical spend. Per-member per-location LTV math sits above this layer for multi-banner multi-location operators and adds: a data-source catalog that joins loyalty data to operator POS (Square POS, Toast POS, Clover, Lightspeed, Aloha, Revel, TouchBistro, ShopKeep, Vend), ecommerce (Shopify, BigCommerce, WooCommerce, Magento, Squarespace Commerce, Wix), CRM (Salesforce, HubSpot, Microsoft Dynamics 365, Zoho, Pipedrive), lifecycle (Klaviyo, Iterable, Braze, Customer.io), subscription (Recharge, Bold Subscriptions, Stripe Billing), reservation (OpenTable, Resy, Tock, SevenRooms), and booking (Square Appointments, Vagaro, MindBody, Booker) data; an LTV class taxonomy covering historical (sum of past purchases), predicted (Bayesian forward-looking), cohort (per-acquisition-cohort), per-channel, per-location, per-banner cross-store portfolio, per-tier (Bronze, Silver, Gold, Platinum), per-RFM-cohort, per-product-category, and per-membership-duration variants; a math-method engine choosing across naive historical, cohort-based average, BTYD (Buy-Till-You-Die) family (BG/NBD, Pareto/NBD, MBG/NBD, BTYD with covariates), Gamma-Gamma monetary value, hierarchical Bayesian (PyMC, Stan, NumPyro, bambi, brms in R, RStan with arviz diagnostics and 95% posterior credible intervals plus Rhat convergence checks), survival analysis (Cox PH, Kaplan-Meier, Weibull AFT, Random Survival Forest, DeepSurv, DeepHit, NN-survival), gradient-boosted trees (XGBoost, LightGBM, CatBoost, NGBoost probabilistic, EBM from InterpretML), neural networks (MLP, RNN, LSTM, Transformer with attention), customer-equity Markov models (Hidden Markov, Continuous-Time Markov Chain), per-location attribution (cross-banner, cross-channel, deterministic and probabilistic identity), and per-attribution-window (30-day, 60-day, 90-day, 180-day, 365-day, 730-day, 1825-day five-year); a cold-start engine using first-purchase priors (per-vertical, per-banner, per-channel, per-Census-ACS demographic, per-Claritas-PRIZM segment), look-alike imputation (K-NN, deep look-alike, Bayesian Optimal Transport), Bayesian shrinkage to portfolio mean, and a 30-day minimum observation window; a refresh-cadence engine (daily for high-tier members, at-risk segments, recent purchasers, and fraud signals; weekly for active members; monthly for the all-member baseline; quarterly for deep recompute with model retraining, hyperparameter tuning, cross-validation, and temporal holdout; event-driven on purchase, return, tier change, complaint resolution, enrollment, data correction, and DSAR); cross-banner handling tying per-banner identity stitching to a per-portfolio member ID with cross-banner earn-redemption and per-banner brand protection under FDD Item 12 territorial rules; anomaly detection on per-member z-score (versus cohort and versus historical self), sudden-spike fraud signals, sudden-drop churn risk, and per-cohort drift; an NBA next-best-action handoff to the loyalty-journey decisioning siblings; a per-member compliance gate (covered below); and an audit trail to operator-controlled WORM storage at per-statute retention windows.

What are the operationally distinctive compliance anchors for loyalty LTV math, and how does the per-member compliance gate cover them?

The per-member compliance gate runs as operator-counsel-approved policy-as-code (OPA Rego, AWS Cedar, Casbin, Cerbos, or Oso). Three anchors are operationally distinctive for loyalty programs that use LTV to drive tier assignment, offer eligibility, or pricing personalization. Anchor 1 — Robinson-Patman Act price discrimination (15 USC 13): loyalty tiers that effectively grant Tier-A buyers materially different per-unit pricing than Tier-B buyers across reseller-style relationships can constitute Robinson-Patman exposure; FTC reinvigorated Robinson-Patman enforcement activity in 2024-2025 against rebate and loyalty mechanics in alcohol distribution, grocery, and pharmacy distribution; programs that grant tier-discounts to business buyers should include a functional-equivalent-availability defense, FTC Section 13(a) defenses, and per-state-equivalent law analysis. Anchor 2 — ECOA Reg B disparate impact (12 CFR 1002): LTV cohort assignment cannot use or proxy for protected class (race, color, religion, national origin, sex including sexual orientation and gender identity, marital status, age, receipt of public assistance, good-faith exercise of Consumer Credit Protection Act rights); disparate-impact analysis is required when ZIP code, surname, or other proxies correlate with protected class; ECOA enforcement extends to non-credit loyalty contexts when LTV-driven offers materially affect credit-adjacent decisions including BNPL eligibility, store-card pre-approval, and subscription-credit terms; Fair Housing Act and HUD v. Facebook (2022) apply when LTV is used for housing-adjacent service (furniture and appliance rent-to-own, home-services scheduling). Anchor 3 — EU AI Act Article 22 plus GDPR Article 22 automated-decisionmaking: when LTV drives automated offer-eligibility, tier-assignment, price personalization, credit decisioning, housing-adjacent decisioning, employment-related decisioning, or scoring-and-ranking that produces significant effects on the data subject, human-review mechanism, right to explanation, right to contest, and right not to be subject to solely automated decisionmaking apply (absent statutory exception). Beyond the three anchors, the gate also covers FTC fake-review rule (16 CFR Part 465) loyalty-incentive disclosure, FTC Negative Option ROSCA and 16 CFR Part 425 when paid loyalty tiers auto-renew, FTC Endorsement Guides 2023 when influencers participate, FTC AI-disclosure, FTC Made in USA, FTC Green Guides, FTC MAP, FTC Health Products Compliance Guide, FTC Section 5, CFPB UDAAP, HIPAA (no acknowledge-as-patient, no PHI in LTV cohort export, HIPAA marketing authorization 164.508), FCRA where LTV influences credit decisioning, GLBA, CCPA/CPRA sensitive personal information and right to opt out of automated decisionmaking, GDPR Articles 6/7/9/17/22, LGPD, DPDP, PIPEDA, COPPA where under-13 loyalty programs apply, FDD Item 12 territorial protection (cross-banner loyalty redemption cannot violate franchisee territorial protection), FDD Item 17 and Item 19, state FRR, state UDTPA, Lanham Act, FINRA Rule 2210 for financial-services loyalty, SEC Reg S-K, state bar advertising for legal-services loyalty, state professional licensing, FDA DSHEA for supplement loyalty, FDA OPDP for Rx-drug loyalty, DEA Schedule II-V, cannabis state board loyalty rules (Massachusetts CCC, California DCC, Colorado MED, Oregon OLCC, Washington WSLCB, Nevada CCB, Illinois CCSL, New Jersey CRC, New York OCM, Arizona DHS, Alaska AMCO, Maine OMP) where most cannabis boards limit or prohibit loyalty programs to prevent inducement, alcohol TABC/CalABC/SLA and DISCUS Code tied-house prohibitions which limit producer/distributor incentives at retailer level, tobacco FDA prohibition in most states, state lottery, EU AI Act Articles 50/13/14/15/22, EU Digital Services Act Article 30, EU Digital Markets Act, WCAG 2.2 AA, ARIA, EAA EN 301 549, Section 508, ADA Title III, Illinois BIPA when facial-recognition loyalty applies, Texas CUBI, Washington MHMDA when health-loyalty applies, NIST AI RMF, ISO 42001, ISO 27001, and SOC 2 Type II.

How does the math-method engine choose between BTYD, hierarchical Bayesian, survival, GBT, and neural-net approaches?

Method selection is operator-data-science-team-aligned and depends on the operator data shape, vertical, and operator-finance-team validation requirements. BTYD family (BG/NBD, Pareto/NBD, MBG/NBD via the lifetimes Python library or PyMC-Marketing) suits non-contractual retail and CPG patterns where customers buy and become inactive without churn signals. Gamma-Gamma monetary-value extension handles the spend-per-transaction side once frequency is modeled. Hierarchical Bayesian models on PyMC, Stan, NumPyro, bambi, or brms in R suit franchise systems where per-location partial pooling on portfolio priors stabilizes small-sample locations. Survival analysis (Cox PH, Kaplan-Meier, Weibull AFT, Random Survival Forest, DeepSurv, DeepHit from scikit-survival or lifelines) suits subscription-style loyalty (Recharge, Bold Subscriptions, Stripe Billing) where time-to-churn is the right modeling target. Gradient-boosted trees (XGBoost, LightGBM, CatBoost, NGBoost for probabilistic predictions, EBM from InterpretML for explainability) suit large-feature operator data warehouses where the operator data-science team can deliver feature engineering at scale. Neural networks (MLP, RNN, LSTM, Transformer with attention) suit very-large-data operators with deep transactional history. Customer-equity Markov models (Hidden Markov, Continuous-Time Markov Chain) suit tier-transition modeling where the operator wants explicit transition probability matrices. The selection logic lives in the operator repo and is calibrated against an operator-data-science-team-graded holdout. Per-location attribution wraps the method choice: cross-banner identity stitching, cross-channel matching, deterministic identity resolution, and probabilistic identity resolution feed identity reconciliation; per-attribution-window (30-day through 1,825-day) is operator-counsel-policy depending on how long the operator can defend the underlying data inputs.

How does cold-start, refresh-cadence, cross-banner handling, and anomaly detection fit together?

Cold start runs first-purchase priors with per-vertical, per-banner, per-channel, per-Census-ACS demographic, and per-Claritas-PRIZM segment defaults; look-alike imputation uses K-nearest-neighbor, deep look-alike encoders, or Bayesian Optimal Transport methods; Bayesian shrinkage to portfolio mean uses hierarchical priors with partial pooling; a 30-day minimum observation window prevents premature classification. Refresh cadence runs daily for high-tier members (Platinum, Gold), at-risk segments, recent purchasers within 7 days, and fraud-signal flags; weekly for active members with purchase within 30 days; monthly for the all-member baseline; quarterly for deep recompute with model retraining, hyperparameter tuning, cross-validation, and temporal holdout; event-driven on purchase, return, tier change, complaint resolution, loyalty program enrollment, data correction, and DSAR request. Refresh feedback closes with realized-vs-predicted LTV, realized-vs-predicted churn, realized-vs-predicted tier-transition, pattern learning, multi-arm bandit regret tracking, and recalibration. Cross-banner handling stitches per-banner identity into a per-portfolio member ID, governs cross-banner earn and redemption, and respects FDD Item 12 territorial protection (cross-banner loyalty redemption cannot violate franchisee territorial protection rules). Anomaly detection runs per-member z-score (versus cohort and versus historical self), sudden-spike fraud signals, sudden-drop churn risk, and per-cohort drift; outliers feed the operator cohort-z-score sibling skill and the operator anomaly-detection agent. The NBA next-best-action handoff routes to the loyalty-journey decisioning siblings (tier-transition timing, cross-location earn-redeem, offer dedup, welcome and milestone and dormant and win-back journey decisioning) plus customer-data-graph consumption, per-channel earn ROI attribution, cross-location earn-redemption eligibility per autonomy profile, email/SMS/social sibling triggers, and the per-location loyalty-dashboard signal handoff to the rollup-reporting agent.

What does the audit trail look like, and how does this skill hand off to the broader operator swarm?

Every LTV computation persists to operator-controlled WORM storage (AWS S3 Object Lock, Google Cloud Storage retention, Azure Blob immutable, or Snowflake Time Travel) with the member-ID, banner, location, data-source snapshot (loyalty, POS, ecommerce, CRM, lifecycle, subscription, reservation, booking), LTV-class snapshot, math-method snapshot per method run, per-location attribution snapshot, per-attribution-window snapshot, LTV confidence-tier snapshot, SHAP/LIME/counterfactual/anchor explainability snapshot, cold-start prior snapshot, refresh-cadence snapshot, FBC feedback-loop snapshot, cross-banner identity-stitching snapshot, FDD Item 12 territorial snapshot, anomaly-detection snapshot, NBA-handoff snapshot, and full compliance-gate snapshot per framework. Retention windows are operator-counsel-policy and typically include IRS 7 years for tax records, FTC substantiation 7 years, HIPAA 6 years where applicable, SEC 3 years where the operator is regulated, FINRA 3 years where applicable, Illinois BIPA biometric retention 3 years where facial-recognition loyalty applies, plus per-state retention. End-to-end replay rewinds any past LTV computation to its source snapshots and re-runs with the current LTV-class, math-method, cold-start, refresh, cross-banner, anomaly, NBA, and compliance versions so the operator data-science and counsel teams can audit changes. Cross-skill handoffs route into the loyalty-journey agent (tier-transition timing, cross-location earn-redeem, offer dedup, welcome and milestone and dormant and win-back decisioning), the customer-data-graph build-pillar, the cross-touchpoint identity resolution build-pillar, the deterministic-probabilistic hybrid identity resolution build-pillar, the runtime-readable behavioral cohorts build-pillar, the versioned customer history DSAR build-pillar, the versioned-history regulatory defense build-pillar, the multi-source attribution-preserving lead ingestion build-pillar, the per-location multi-model attribution build-pillar, the CRM record creation build-pillar, the per-location dynamic content build-pillar (LTV quintile drives dynamic content), the lifecycle email-SMS build-pillar, the multi-location SMS broadcast build-pillar, the per-location SMS template library build-pillar, the callback-schedule-link build-pillar (LTV tier drives callback rep assignment), the per-location missed-call CRM creation and callback workflow build-pillar, the event tie-in drafting build-pillar, the weather and seasonality patterns build-pillar, the local-context change-events build-pillar, the CS agent assist build-pillar (LTV tier drives CS rep prioritization), the review-response drafting build-pillar, the per-location per-cohort two-sigma anomaly detection build-pillar (LTV outliers feed cohort z-score), the alert deduplication build-pillar, the per-location post-drafting build-pillar, the ad-performance feedback-loop build-pillar (LTV quintile drives audience targeting), the tiered pre-filter deterministic gates build-pillar, the marketing-content LLM-as-judge build-pillar, the marketing-AI autonomy-profile configuration build-pillar, the per-jurisdiction compliance for multi-state franchise build-pillar, the routing audit trail build-pillar, the borderline routing skill, the five-destination routing skill, the FBC override learning skill, the multi-dimensional threshold routing skill, the brand-voice management skill, the forbidden-phrase library skill, the claims-allowlist substantiation skill, the changelog feed ingestion build-pillar (loyalty platform changelog ingestion), and the foot-traffic integration build-pillar (loyalty-tier-aware foot-traffic attribution).

Engage Completions

Completions builds and operates the per-member LTV math bundle on the loyalty-journey agent end-to-end. Operator owns the data-source catalog, the LTV class taxonomy, the math-method engine, the cold-start and refresh engines, the cross-banner handling, the anomaly detection, the NBA handoff, the compliance overlay rule library, the orchestration code, the LLM prompts, and the audit trail. Operator can in-house at any time.