Measure-improve swarm · Offline-attribution-intelligence agent · Per-location-attribution-models skill · Build pillar · Published July 15, 2026
How to build per-location multi-model attribution at multi-store scale
Multi-store operators running attribution per location work above a strong web-analytics + MMP + MTA + MMM + foot-traffic + identity- resolution + ML-platform + policy-as-code primitives layer (Google Analytics 4 + Adobe Analytics + Mixpanel + Amplitude + Heap for web analytics; Singular + AppsFlyer + Adjust + Branch + Kochava for mobile-measurement-partner attribution; Rockerbox + Northbeam + Triple Whale + Hyros + Wicked Reports + Measured for multi- touch attribution; Analytic Edge + Marketing Evolution + Nielsen MTA + Neustar MarketShare for commercial MMM; Placer.ai + SafeGraph + Foursquare + Veraset + Cuebiq + Near + Unacast for foot-traffic; Tealium + Segment + mParticle + RudderStack for identity resolution; Vertex AI + AWS SageMaker + Azure ML + Databricks ML for ML platforms — each vendor ships sophisticated primitives that the orchestration sits above). The orchestration that sits above those primitives — a per-location per-touchpoint catalog covering 20+ touchpoint types, a per-location per- attribution-model stack across rule-based, statistical, machine- learning, and causal-inference families, a per-location per- model reconciliation layer with operator-counsel-and-finance- approved rules, a multi-LLM pre-publish substantiation check, a feedback loop comparing realized vs attributed revenue and incrementality, and a per-model compliance gate that ties decisions to FTC substantiation + FTC MARS + FDD Item 19 + SEC Reg S-K + SOX 302/404 + causal-vs-correlation anchors — is operator-side architecture. Account-level data-driven attribution scores treat the operator as a single attribution surface; per- location multi-model attribution treats every store as a modeling unit with its own touchpoint mix, model stack, and reconciliation. This guide explains how to architect the per- location-attribution-models skill on the offline-attribution- intelligence agent end-to-end.
What you will build
- A per-location per-touchpoint catalog covering 20+ touchpoint types: paid-search click, paid-social click, paid-display impression, paid-display click, organic-search click, organic-social engagement, direct visit, referral visit, email open, email click, SMS click, push click, app deep link, call-tracking call, direct-mail PURL, direct-mail QR, direct-mail promo code, foot-traffic visit, in-store receipt, loyalty redemption.
- A per-location per-attribution-model stackwith first-touch, last-touch, last-non-direct, linear equal- credit, time-decay half-life, position-based 40/20/40 U-shape, W-shape, Z-shape, custom rule-based, data-driven Shapley- value, Markov-chain removal-effect, survival-analysis, Bayesian causal inference, incrementality control-vs-holdout, geo-experiment incrementality, ghost-bidding incrementality, MMM coefficient attribution, MTA, unified MMM-MTA, causal- uplift, and multi-LLM ensemble models (operator-chosen across the catalog and across LLM vendors).
- A per-location per-model reconciliation layer with operator-counsel-and-finance-approved rules: revenue- budget conservation, channel-coefficient bound, saturation- adstock correction, halo-cannibalization correction, cross- model revenue-sum tolerance, cross-model revenue-deviation flag, cross-model incrementality floor and ceiling.
- A multi-LLM pre-publish substantiation checkensembling multiple vendor LLM APIs for attribution claim, revenue claim, incrementality claim, causal-vs-correlation disambiguation, FTC, FDD, FINRA, CFPB, and SEC cross-check with confidence scoring, self-consistency cross-check, and chain-of-thought extraction.
- A per-model compliance gate anchored on FTC substantiation doctrine + FTC MARS multi-location claim consistency + FDD Item 19 Financial Performance Representations + SEC Reg S-K Item 303 MD&A + SEC Reg G non-GAAP financial measures + Sarbanes-Oxley Section 302 CEO/CFO certification + Section 404 internal control attestation + causal-vs-correlation disambiguation, extended to FTC AI disclosure + FTC Endorsement Guides 2024 + FINRA Rule 2210 + CFPB UDAAP + CCPA/CPRA + GDPR + PIPEDA + CASL + LGPD + DPDP + EU AI Act Articles 13/14/15 + NIST AI RMF + ISO 42001 + ISO 27001 + SOC 2 Type II via policy-as-code (OPA Rego, AWS Cedar, Casbin, Cerbos, Oso) that operator counsel reviews.
- A feedback loop comparing realized vs attributed revenue (MAPE, WAPE), realized vs attributed incrementality (against control-vs-holdout and geo-experiment validation), with MMM coefficient, Markov removal-effect, Shapley value, Bayesian prior, causal-uplift, time-decay, position-based weight, saturation-adstock, and halo- cannibalization recalibration plus model-drift detection and retraining-trigger logic.
- Cross-skill handoffs and an audit trail to siblings on the offline-attribution-intelligence agent and broader swarm, with audit trail to operator-controlled WORM storage at per-statute retention windows operator counsel sets.
Where the orchestration above web-analytics, MMP, MTA, MMM, and ML-platform primitives compounds at multi-store scale
The vendor primitives are strong. Web-analytics vendors expose per-account event capture. MMP vendors handle mobile-touchpoint attribution. MTA vendors compute cross-channel attribution. Commercial MMM vendors handle account-rolled-up MMM. Foot- traffic vendors expose per-location visit signal. Identity- resolution vendors stitch cross-touchpoint identity. ML platforms handle model lifecycle. The orchestration above those primitives is what compounds at multi-store attribution scale.
The first operationally distinctive constraint is FTC substantiation doctrine plus FTC MARS multi-location claim consistency. Attribution-derived ROAS, revenue-attribution, and incrementality figures must be substantiable. When per- location attribution outputs differ across locations for the same operator brand, FTC MARS applies. The per-model gate ties every published number to the underlying attribution run, model parameters, and reconciliation evidence.
The second distinctive constraint is FDD Item 19 Financial Performance Representations. When per-location attribution outputs are shared with the franchisee council or prospective franchisees, Item 19 governs the representation (basis, time period, geographic scope, substantiation). The gate routes franchisee-visible attribution outputs to the FDD Item 19 workflow before release.
The third distinctive constraint is SEC Regulation S-K Item 303 (MD&A) + SEC Regulation G non-GAAP financial measures when attribution-derived metrics surface in public filings, earnings releases, or investor decks. Reg G requires reconciliation to the most directly comparable GAAP measure. SEC Regulation S-K Item 506 forward-looking-statement framework applies when attribution drives forward-looking ROAS or incrementality projections. The gate routes publicly- disclosed outputs to securities-counsel review.
The fourth distinctive constraint is Sarbanes-Oxley Section 302 CEO/CFO certification + Section 404 internal control attestation. Attribution outputs that influence revenue allocation across channels or locations are part of the financial-reporting control surface; SOX 302 requires certification and SOX 404 requires internal-control attestation. The gate ties every attribution run to the operator-finance-team-approved control evidence record.
The fifth distinctive constraint is causal-vs-correlation disambiguation. Attribution models produce correlational estimates whose causal interpretation depends on identification assumptions (no-confounding, stable-unit- treatment-value, conditional ignorability). FTC Reasonable- Basis Doctrine and standard scientific practice both require that claims framed as causal be backed by causal- identification evidence. The gate enforces a causal-vs- correlation disclosure on every published output.
Beyond the five anchors, the gate also covers FTC AI disclosure when AI drives the attribution; FTC Endorsement Guides 2024; FINRA Rule 2210 when investment-grade operators publish attribution-derived performance; CFPB UDAAP when attribution touches consumer-finance decisioning; CCPA/CPRA + GDPR + PIPEDA + CASL + LGPD + DPDP when attribution joins customer identity; EU AI Act Articles 13/14/15 when AI- driven attribution drives automated decisioning; NIST AI RMF + ISO 42001 + ISO 27001 + SOC 2 Type II for the attribution-governance infrastructure. The gate is policy- as-code; operator counsel reviews rule updates.
The real ecosystem the orchestration sits above
Web-analytics, MMP, and MTA primitives
Google Analytics 4, Adobe Analytics, Mixpanel, Amplitude, Heap for web analytics; Singular, AppsFlyer, Adjust, Branch, Kochava for mobile-measurement-partner attribution; Rockerbox, Northbeam, Triple Whale, Hyros, Wicked Reports, Measured for multi-touch attribution. Strong primitives. The per-location per-touchpoint catalog + per-attribution- model stack layers sit above this layer.
MMM, foot-traffic, and identity-resolution primitives
Analytic Edge, Marketing Evolution, Nielsen MTA, Neustar MarketShare for commercial MMM; Placer.ai, SafeGraph, Foursquare, Veraset, Cuebiq, Near, Unacast for foot-traffic; Tealium, Segment, mParticle, RudderStack for identity resolution. Strong primitives. The per-location per-model reconciliation + per-model compliance gate layers compose them under operator-counsel-and-finance-reviewed governance.
ML-platform and model-explainability primitives
Vertex AI, AWS SageMaker, Azure ML, Databricks ML for ML platforms; MLflow, Weights & Biases, Comet ML for experiment tracking; SHAP, LIME, Captum, InterpretML for explainability. Strong primitives. The per-location per- attribution-model stack + per-model reconciliation + feedback-loop layers sit above this layer.
Policy-as-code, WORM-storage, and compliance-tooling primitives
OPA Rego, AWS Cedar, Casbin, Cerbos, Oso for policy-as- code; AWS S3 Object Lock, GCS retention, Azure Blob immutable, Snowflake Time Travel for WORM storage; Hyperproof, Drata, Vanta, Thoropass for SOC 2 / ISO control evidence; OneTrust, TrustArc, Ketch for privacy program tooling. Strong primitives. The per-model compliance overlay coordinates them via the policy-as-code gate that operator counsel reviews.
How the architecture is built
- Per-location per-touchpoint catalog.Subscribe to ad-platform APIs, MMP attribution events, web- analytics events, lifecycle vendor events, call-tracking events, vanity-URL/vanity-QR/promo-code registries, foot- traffic vendor APIs, receipt-joining handoffs, and loyalty- management handoffs. Canonicalize to the per-location grain and join to the operator customer canonical-ID.
- Per-location per-attribution-model stack.Run every operator-chosen model on the operator-chosen ML platform. Emit per-location attribution tables with confidence tier and explainability per model.
- Per-location per-model reconciliation. Apply operator-counsel-and-finance-approved rules (revenue-budget conservation, channel-coefficient bound, saturation-adstock correction, halo-cannibalization correction, cross-model revenue-sum tolerance, cross-model revenue-deviation flag, cross-model incrementality floor and ceiling).
- Multi-LLM pre-publish substantiation check.Ensemble multiple vendor LLM APIs. Cross-check attribution claim substantiation, revenue claim substantiation, incrementality claim substantiation, causal-vs-correlation disambiguation, and per-statute compliance. Run self- consistency checks. Extract chain-of-thought to the audit trail.
- Per-model compliance gate. Express the gate as policy-as-code on OPA Rego, AWS Cedar, Casbin, Cerbos, or Oso. Encode the five distinctive anchors (FTC substantiation + FTC MARS, FDD Item 19, SEC Reg S-K + Reg G, SOX 302/404, causal-vs-correlation) plus the broader compliance surface. Operator counsel reviews every rule update.
- Feedback loop. Compare realized vs attributed revenue (MAPE, WAPE). Compare realized vs attributed incrementality against control-vs-holdout validation and geo-experiment validation. Recalibrate MMM coefficients, Markov removal effects, Shapley values, Bayesian priors, causal-uplift estimates, time-decay half- life, position-based weights, saturation-adstock, and halo- cannibalization decompositions. Detect model drift. Trigger retraining.
- Cross-skill handoffs. Hand off to siblings on the offline-attribution-intelligence agent and broader swarm.
- Audit trail. Emit a per-attribution-run canonical audit record to operator-controlled WORM storage with per-statute retention windows operator counsel sets (IRS 7yr, FTC 7yr, SEC 6yr, SOX 7yr).
Frequently asked
What does per-location multi-model attribution do that an account-level data-driven attribution score does not?
Web analytics vendors (Google Analytics 4, Adobe Analytics, Mixpanel, Amplitude, Heap) ship strong primitives for per-account event capture and data-driven attribution. Mobile-measurement-partner (MMP) vendors (Singular, AppsFlyer, Adjust, Branch, Kochava) ship strong primitives for mobile-touchpoint attribution. Multi-touch attribution (MTA) vendors (Rockerbox, Northbeam, Triple Whale, Hyros, Wicked Reports, Measured) ship strong primitives for cross-channel attribution. Commercial MMM vendors (Analytic Edge, Marketing Evolution, Nielsen MTA, Neustar MarketShare) ship strong primitives for account-rolled-up MMM. Foot-traffic vendors (Placer.ai, SafeGraph, Foursquare, Veraset, Cuebiq, Near, Unacast) ship strong primitives for per-location visit signal. Identity-resolution vendors (Tealium, Segment, mParticle, RudderStack) ship strong primitives for cross-touchpoint identity stitching. Per-location multi-model attribution sits above this layer for multi-store operators running attribution per location, and adds: a per-location per-touchpoint catalog covering 20+ touchpoint types (paid-search click, paid-social click, paid-display impression, paid-display click, organic-search click, organic-social engagement, direct visit, referral visit, email open, email click, SMS click, push click, app deep link, call-tracking call, direct-mail PURL, direct-mail QR, direct-mail promo code, foot-traffic visit, in-store receipt, loyalty redemption); a per-location per-attribution-model stack with first-touch, last-touch, last-non-direct touch, linear equal-credit, time-decay half-life, position-based 40/20/40 U-shape, W-shape, Z-shape full-path, custom rule-based, data-driven Shapley-value, Markov-chain removal-effect, survival-analysis, Bayesian causal inference, incrementality control-vs-holdout, geo-experiment incrementality, ghost-bidding incrementality, MMM coefficient attribution, multi-touch attribution, unified MMM-MTA, causal-uplift, and multi-LLM ensemble models (operator chooses across the catalog and across LLM vendors — OpenAI, Anthropic, Google, Mistral, Cohere, Meta), with per-location attribution confidence tier and explainability; a per-location per-model reconciliation layer that applies operator-counsel-and-finance-approved reconciliation rules (revenue-budget conservation, channel-coefficient bound, saturation-adstock correction, halo-cannibalization correction, cross-model revenue-sum tolerance, cross-model revenue-deviation flag, cross-model incrementality floor, cross-model incrementality ceiling); a multi-LLM pre-publish check that cross-checks attribution claim substantiation, revenue claim substantiation, incrementality claim substantiation, causal-vs-correlation disambiguation, and per-statute compliance; a per-model compliance gate (covered in the next answer); a feedback loop that compares realized vs attributed revenue (MAPE, WAPE) and realized vs attributed incrementality against control-vs-holdout validation and geo-experiment validation, recalibrating MMM coefficients, Markov removal effects, Shapley values, Bayesian priors, causal-uplift estimates, time-decay half-life, position-based weights, saturation-adstock, and halo-cannibalization decompositions; and a per-attribution-run audit record to operator-controlled WORM storage at per-statute retention windows.
What are the operationally distinctive compliance anchors for per-location multi-model attribution, and how does the per-model compliance gate cover them?
Five anchors sit at the operational center of multi-store attribution that off-the-shelf attribution compliance overlays often miss. Anchor 1 — FTC substantiation doctrine + FTC MARS multi-location claim consistency. Attribution-derived ROAS, revenue-attribution, and incrementality figures surfaced as advertising claims, sales-collateral claims, or franchisee-recruitment claims must be substantiable under the FTC Reasonable-Basis Doctrine (Pfizer 1972). When per-location attribution outputs differ across locations for the same operator brand, FTC MARS multi-location claim consistency applies. The per-model gate ties every published number to the underlying attribution run, model parameters, and reconciliation evidence so counsel has the substantiation record. Anchor 2 — FDD Item 19 Financial Performance Representations. When per-location attribution outputs are shared with the franchisee council or prospective franchisees, Item 19 governs how those representations can be made (basis, time period, geographic scope, substantiation). The per-model gate routes franchisee-visible attribution outputs to the FDD Item 19 workflow before release. Anchor 3 — SEC Regulation S-K Item 303 (MD&A) + SEC Regulation G non-GAAP financial measures. When attribution-derived metrics surface in public filings, earnings releases, or investor decks, Reg S-K Item 303 governs the discussion and analysis of financial condition, and Reg G governs non-GAAP financial measure disclosure requiring reconciliation to the most directly comparable GAAP measure. SEC Regulation S-K Item 506 forward-looking-statement framework applies when attribution drives forward-looking ROAS or incrementality projections. The per-model gate routes publicly-disclosed attribution outputs to securities-counsel review. Anchor 4 — Sarbanes-Oxley Section 302 CEO/CFO certification + Section 404 internal control attestation. Attribution outputs that influence revenue allocation across channels or locations are part of the financial-reporting control surface; SOX 302 requires CEO/CFO certification and SOX 404 requires internal-control attestation. The per-model gate ties every attribution run to the operator-finance-team-approved control evidence record. Anchor 5 — causal-vs-correlation disambiguation. Attribution models produce correlational estimates whose causal interpretation depends on identification assumptions (no-confounding, stable-unit-treatment-value, conditional ignorability). FTC Reasonable-Basis Doctrine and standard scientific practice both require that claims framed as causal be backed by causal-identification evidence. The per-model gate enforces a causal-vs-correlation disclosure on every published output and pairs MTA/MMM-derived estimates with incrementality-test results when available. Beyond the five anchors, the per-model gate also covers FTC AI disclosure when AI drives the attribution, FTC Endorsement Guides 2024 when attribution-derived endorsements surface; FINRA Rule 2210 when investment-grade operators publish attribution-derived performance; CFPB UDAAP when attribution touches consumer-finance decisioning; CCPA/CPRA + GDPR + PIPEDA + CASL + LGPD + DPDP when attribution joins customer identity; EU AI Act Articles 13/14/15 when AI-driven attribution drives automated decisioning; NIST AI RMF + ISO 42001 + ISO 27001 + SOC 2 Type II for the attribution-governance infrastructure. The gate is policy-as-code on OPA Rego, AWS Cedar, Casbin, Cerbos, or Oso, with operator counsel reviewing rule updates.
How do the per-location per-touchpoint catalog, per-location per-attribution-model stack, and per-model reconciliation layer actually work?
The per-location per-touchpoint catalog ingests 20+ touchpoint types per location. Paid touchpoints (paid-search click, paid-social click, paid-display impression, paid-display click) come from ad-platform APIs (Google Ads, Meta, TikTok, Amazon Ads, Microsoft Ads, Pinterest, Snap, Reddit, LinkedIn) plus MMP attribution events (Singular, AppsFlyer, Adjust, Branch, Kochava). Organic touchpoints (organic-search click, organic-social engagement) come from Google Search Console + Bing Webmaster + organic-social platforms. Direct and referral visits come from the operator web analytics layer (Google Analytics 4, Adobe Analytics, Mixpanel, Amplitude, Heap). Lifecycle touchpoints (email open, email click, SMS click, push click, app deep link) come from lifecycle vendors (Klaviyo, Iterable, Braze, Customer.io, Twilio, Bandwidth, MessageBird). Call-tracking touchpoints come from per-location call-tracking integration. Direct-mail touchpoints (PURL, QR, promo code) come from operator-controlled vanity-URL, vanity-QR, and promo-code registries. Offline touchpoints (foot-traffic visit, in-store receipt, loyalty redemption) come from the foot-traffic-integration, receipt-joining, and loyalty-management agent siblings. Each touchpoint event is canonicalized to the per-location grain and joined to the operator-maintained customer canonical-ID. The per-location per-attribution-model stack runs every model the operator chooses to operate (first-touch, last-touch, last-non-direct, linear, time-decay, position-based 40/20/40 U-shape, W-shape, Z-shape, custom rule-based, data-driven Shapley, Markov chain removal-effect, survival analysis, Bayesian causal inference, incrementality control-vs-holdout, geo-experiment incrementality, ghost-bidding incrementality, MMM coefficient attribution, MTA, unified MMM-MTA, causal-uplift, multi-LLM ensemble) on operator-chosen ML platform (Vertex AI, AWS SageMaker, Azure ML, Databricks ML). Each model emits a per-location attribution table with confidence tier and explainability. The per-model reconciliation layer applies operator-counsel-and-finance-approved rules: revenue-budget conservation (every attributed dollar must trace to actual revenue), channel-coefficient bound (per-channel coefficients stay within operator-counsel-approved bounds), saturation-adstock correction (apply per-channel saturation curve and adstock decay per the MMM handoff), halo-cannibalization correction (apply cross-channel halo/cannibalization decomposition), cross-model revenue-sum tolerance (model-output sums fall within operator-counsel-approved tolerance), cross-model revenue-deviation flag (flag deviations beyond tolerance for counsel review), cross-model incrementality floor and ceiling (operator-counsel-approved incrementality range).
How do the multi-LLM pre-publish check, feedback loop, and cross-skill handoffs coordinate with the rest of the swarm?
The multi-LLM pre-publish check ensembles multiple vendor LLM APIs (operator chooses across OpenAI, Anthropic, Google, Mistral, Cohere, Meta) to cross-check attribution claim substantiation, revenue claim substantiation, incrementality claim substantiation, causal-vs-correlation disambiguation, FTC substantiation, FTC MARS, FDD Item 19 representation, FINRA 2210, CFPB UDAAP, and SEC Reg S-K. Self-consistency cross-check and chain-of-thought extraction populate the audit trail. The feedback loop compares realized vs attributed revenue via MAPE and WAPE, compares realized vs attributed incrementality against control-vs-holdout validation and geo-experiment validation, and recalibrates MMM coefficients, Markov removal effects, Shapley values, Bayesian priors, causal-uplift estimates, time-decay half-life, position-based weights, saturation-adstock, and halo-cannibalization decompositions. Pattern learning, emerging-channel detection, model drift detection, and retraining-trigger logic carry the operator-counsel-reviewed governance through the model lifecycle. The skill hands off to siblings on the offline-attribution-intelligence agent (multi-vendor receipt joining, foot-traffic integration, multi-vendor call tracking, per-location cross-channel attribution rollup, per-location MMM, per-location MMM-driven budget recommendation engine, root-cause attribution sketch, multi-source attribution-preserving lead ingestion, cross-touchpoint identity resolution, deterministic-probabilistic hybrid identity resolution) and across the broader swarm (marketing forecasting, cross-stream correlation, routing audit trails, versioned history for regulatory defense, per-location metric ingestion, marketing-stack integration health, per-location rollup reporting, customer data graph, brand-voice management).
What does Completions report on a Tier 3 engagement that covers per-location multi-model attribution?
Tier 3 engagements report against a pre-engagement baseline that the Tier 1 assessment establishes for the operator stack. The reporting cycle covers six workstreams: (1) per-location per-touchpoint catalog coverage observed across the 20+ touchpoint type surface, with per-source ingestion completeness and identity-resolution coverage reported; (2) per-location per-attribution-model stack surface observed across the model catalog operator data-science maintains, with per-model goodness-of-fit and per-model confidence diagnostics reported; (3) per-location per-model reconciliation surface observed against operator-counsel-and-finance-approved rules, with per-rule conflict-resolution diagnostics and cross-model revenue-sum tolerance observations reported; (4) multi-LLM pre-publish substantiation check surface observed against operator-labeled holdouts, with per-vendor confidence diagnostics reported; (5) feedback-loop surface observed across realized vs attributed revenue, realized vs attributed incrementality, and per-method recalibration diagnostics, with model-drift detection and retraining-trigger observations reported; (6) per-model compliance gate pass rate observed across FTC substantiation + FTC MARS + FTC AI disclosure + FTC Endorsement Guides 2024 + FDD Item 19 + SEC Reg S-K Item 303 + SEC Reg G + Sarbanes-Oxley Section 302/404 + causal-vs-correlation disambiguation + FINRA Rule 2210 + CFPB UDAAP + CCPA/CPRA + GDPR + EU AI Act Articles 13/14/15 + NIST AI RMF + ISO 42001 + ISO 27001 + SOC 2 Type II scope. Caveats: per-source vendor API rate limits + per-source ingestion completeness + foot-traffic-vendor sample quality + LLM-vendor availability + per-statute retention windows shifting with operator counsel policy + EU AI Act high-risk-system designation updates sit outside Completions control and are reported alongside observed performance; attorney-client privilege on counsel-and-finance-reviewed reconciliation rules, FDD Item 19 disclosure rules, SEC Reg G non-GAAP-measure disclosures, SOX 302/404 internal-control evidence, and substantiation records is preserved through every layer. Completions does not commit to fixed numeric SLAs on touchpoint coverage, model goodness-of-fit, reconciliation accuracy, MAPE/WAPE, or compliance pass rate when those KPIs depend on vendor performance, sample quality, or counsel policy decisions.
Engage Completions
Start with the AI Readiness Assessment (Tier 1, 2-3 weeks). If the operation is ready to absorb the per-location- attribution-models skill on the offline-attribution- intelligence agent, the assessment hands off to the AI Swarm Setup Sprint (Tier 2, 4-8 weeks). If the operation needs ongoing orchestration after Tier 2 hand-off, the skill continues under Fractional CMO with AI Swarm (Tier 3, 6-month minimum, 1-2 days/wk embedded). Operator owns every artifact at every tier. Operator can in-house at any time.