# Meta Attribution & Incrementality Framework v1.0 ## Enterprise Operating System Guide **Purpose:** Govern how Meta conversion credit, customer journeys, causal impact, and commercial value are defined, reconciled, tested, reported, and used for investment decisions. **Companion guides:** Meta Setup & Activation, Measurement, Reach & Frequency, Optimization, Testing & Experimentation, Reporting & Insights, Creative Strategy, and Troubleshooting & Root Cause. **Core rule:** Attribution allocates credit under configured rules. Incrementality estimates outcomes caused by media relative to a credible counterfactual. Neither should be confused with validated revenue, margin, or profit. --- # 1. Enterprise Mandate The Meta Agent must answer: 1. Which event or campaign received credit? 2. Under which attribution setting and interaction type? 3. Which customer journey was recorded? 4. Which business outcome was validated externally? 5. What would likely have happened without the media? 6. Which additional outcomes were caused by treatment? 7. How uncertain is the estimate? 8. What decision follows? 9. Who owns and approves the interpretation? 10. What evidence, QA, registry, and change-control records are required? Every output must identify whether the result is observed, attributed, modelled, experimental, incremental, or commercially validated. --- # 2. Core Principles - Attributed conversions are not incremental conversions. - Credit is not causality. - Recorded paths are not proof of influence. - Exposure before conversion does not prove the exposure created the conversion. - View attribution can capture plausible influence and existing demand. - Platform attribution does not replace CRM, commerce, sales, or finance validation. - Incrementality requires a credible counterfactual. - Statistical evidence must be interpreted with practical business value. - A non-detected lift does not automatically prove no effect. - A positive lift does not guarantee profitable scale. - Do not generalize beyond tested conditions. - Never invent attribution settings, lift, significance, confidence, or sample requirements. An internal measurement masterclass training resource defines incrementality as additional conversions that would not otherwise have occurred and describes Conversion Lift through test and control groups. --- # 3. Evidence Architecture ## Measurement layer Browser, server, app, offline, CRM, messaging, and commerce events establish signals. ## Attribution layer Meta allocates campaign credit according to supported attribution settings and interaction rules. ## Journey layer Meta, analytics, and business systems describe recorded sequences and funnel progression. ## Incrementality layer Controlled or approved causal methods estimate additional outcomes caused by treatment. ## Business-value layer CRM, commerce, sales, booking, subscription, and finance systems validate qualified outcomes, net revenue, margin, profit, and customer value. The layers answer different questions and must be reported separately. --- # 4. Decision Tree 1: Measurement Method Selection ```text What is the question? ↓ Which campaign received credit? → Attribution Which interactions occurred? → Journey analysis Did media cause additional outcomes? → Incrementality Did outcomes create commercial value? → Business systems Where should the next budget go? → Combine all layers plus marginal return ``` --- # 5. Attribution Source Framework ## Meta Ads Manager Use for Meta campaign delivery and attributed outcomes under the selected reporting configuration. ## Website or app analytics Use for site sessions, app behaviour, product exploration, and the approved analytics attribution view. ## CRM and lead systems Use for accepted, qualified, opportunity, closed, and revenue-generating lead stages. ## Commerce and finance Use for valid orders, refunds, cancellations, net revenue, margin, and profit. ## Incrementality study output Use for the scoped causal estimate produced by the approved study. The primary source depends on the business question. Do not choose the largest number. --- # 6. Decision Tree 2: Attribution Source Selection ```text Need delivery and Meta credit? Yes → Meta reporting Need website or app behaviour? Yes → Approved analytics source Need qualified lead or customer? Yes → CRM or commerce Need net revenue, margin, or profit? Yes → Finance-approved source Need causal impact? Yes → Approved incrementality study ``` --- # 7. Attribution Settings Governance Document the active Meta attribution configuration using the exact product labels available at the time of setup. Record: - Ad account - Campaign and ad set - Conversion location - Optimization event - Attribution setting - Eligible interaction types - Relevant windows - Reporting period and timezone - Event source - Modelled or observed limitations - Effective date - Owner and approver Meta settings can change over time. Verify the active interface and official documentation before implementation. Do not rely on a historical default. --- # 8. Click Attribution Framework Click-attributed outcomes are conversions credited after an eligible click under the configured rules. Review: - Click definition - Link and destination continuity - Landing-page arrival - Conversion lag - Duplicate clicks - Event quality - Customer qualification - Existing demand - Cross-device treatment A click may reflect created demand, captured demand, or navigation by an already-intending customer. Attribution alone cannot separate these possibilities. --- # 9. View and Engagement Attribution Framework Where supported, view or engagement-based attribution can assign credit when an eligible exposure or interaction precedes conversion without a credited link click. Review: - Exact interaction definition - Window - Format eligibility - Frequency - Video or engagement treatment - Retargeting concentration - Customer status - Other media - Existing demand - Modelled components Do not report view- or engagement-attributed outcomes as incremental sales. Separate each interaction type where available. --- # 10. Decision Tree 3: Attribution Window Comparison ```text Comparing campaigns or periods? ↓ Same event, attribution setting, interaction types, windows, and timezone? No → Do not compare directly Yes ↓ Customer decision cycle aligned? No → Run sensitivity analysis Yes ↓ Reconcile with CRM or commerce outcomes ``` Window sensitivity explains credit movement. It does not prove causality. --- # 11. Journey and Assist Framework Journey analysis can show recorded touchpoints before a conversion. Use it to explore: - Sequence - Lag - Repeated exposure - Campaign participation - Funnel progression - Cross-channel hypotheses Do not: - Add assists and attributed conversions as unique business outcomes. - Assume every touchpoint influenced the customer. - Treat path position as causal contribution. - Infer unobserved customer thinking. The journey layer should generate hypotheses for testing and inform reporting context. --- # 12. Decision Tree 4: Journey Analysis ```text Need to understand recorded path? ↓ Are interaction definitions and identity coverage known? No → State limitation Yes ↓ Map sequence, lag, and funnel events ↓ Does the pattern imply causality? No → Treat as descriptive evidence ↓ Create testable hypothesis if decision is material ``` --- # 13. Attribution Bias Framework Potential biases include: - Last-touch bias - View-through over-crediting - Retargeting selection bias - Platform self-attribution - Missing cross-device observations - Consent and identity gaps - Existing-customer concentration - Conversion-lag truncation - Channel classification differences - Survivor bias in observed paths The presence of bias does not make attribution useless. It limits the decisions attribution can support. --- # 14. Cross-Source Reconciliation Differences between Meta, analytics, CRM, commerce, and finance can arise from: - Attribution settings - Click, view, or engagement treatment - Conversion definitions - Timezone - Currency - Cross-device identity - Consent - Event latency - Duplicate handling - Refunds and cancellations - Lead qualification - New versus existing customers For every difference, classify the amount as explained, unresolved, or invalid. --- # 15. Decision Tree 5: Attribution Conflict Diagnosis ```text Meta and business source differ ↓ Same outcome definition? No → Align definitions Yes ↓ Same period, timezone, and currency? No → Restate comparison Yes ↓ Check event receipt, duplicates, attribution, identity, lag ↓ Check qualification, refunds, cancellations, customer status ↓ Classify explained, unresolved, or invalid gap ``` --- # 16. Attribution Reporting Standard Every report must state: - Business question - Meta event - Business source of truth - Attribution setting - Interaction types - Windows - Period and timezone - Attributed result - CRM or commerce result - Explained discrepancy - Limitations - Incrementality evidence, if available Use language such as: - "Meta-attributed conversions under the stated setting" - "Recorded journey participation" - "Measured incremental conversions from the approved study" Never use these phrases interchangeably. --- # 17. Decision Tree 6: Attribution Model Use ```text Is the model used for campaign optimization? Yes → Confirm reliable event and qualified-outcome monitoring Is the model used for executive reporting? Yes → Reconcile with business source Is the model used for budget allocation? Yes → Add incrementality and marginal evidence Is the model used as proof of causality? Yes → Reject interpretation and use a causal method ``` --- # 18. Incrementality Fundamentals Incrementality estimates: ```text Observed outcome under treatment minus Estimated outcome without treatment = Incremental effect ``` The "without treatment" outcome is the counterfactual. It must be estimated through a credible design. An internal measurement masterclass training resource explicitly describes a test group with the opportunity to see ads and a control group that does not, with conversion differences used to calculate incremental lift. --- # 19. Decision Tree 7: Attribution or Incrementality ```text Need to know what received credit? → Attribution Need to know what would have happened anyway? → Incrementality Need both operational and strategic view? → Report attribution and incrementality separately No credible counterfactual available? → Use observational evidence and state limitation ``` --- # 20. Counterfactual Framework A credible counterfactual should be: - Eligible for the same outcome - Comparable before treatment - Protected from treatment exposure - Measured consistently - Subject to similar external conditions - Suitable for the selected analytical method Threats include: - Treatment leakage - Cross-channel contamination - Different promotions - Market shocks - Distribution changes - Tracking differences - Customer migration - Noncompliance - Insufficient outcome volume Weak counterfactuals require weaker causal language. --- # 21. Conversion Lift Framework Use Conversion Lift for approved questions about additional conversions caused by Meta treatment. Govern: - Business question - Eligible campaigns - Conversion event - Data source - Test and control assignment - Study period - Conversion lag - Treatment delivery - Contamination - Result uncertainty - Practical significance - Decision and rollout An internal measurement masterclass training resource positions Conversion Lift for measuring incremental conversions and comparing conversion impact between strategies. Do not invent eligibility, minimums, or expected lift. Confirm current requirements in official Meta documentation and the active platform. --- # 22. Brand Lift Framework Brand Lift measures approved brand-perception outcomes rather than sales credit. Potential outcomes can include: - Ad recall - Awareness - Familiarity - Consideration - Preference - Purchase intent Document: - Question - Market and language - Campaign scope - Exposed and control logic - Survey period - Other media - Result and uncertainty - Scope of inference A positive brand result does not automatically prove revenue impact. --- # 23. Decision Tree 8: Lift Study Selection ```text What outcome needs causal evidence? Brand perception → Brand Lift Conversion or qualified action → Conversion Lift Regional sales or traffic → Geo or matched-market design Long-term portfolio effect → Approved modelling plus experiments No eligible study available → Alternative controlled design or observational limitation ``` --- # 24. Randomized Holdout Framework Define: - Unit of assignment - Eligible population - Treatment - Control - Allocation - Primary outcome - Guardrails - Sample and duration logic - Conversion lag - Cross-device treatment - Contamination - Analysis method Treatment assignment or opportunity to see is not the same as confirmed individual exposure. Report the design accurately. --- # 25. Decision Tree 9: Holdout Design ```text Can eligible units be randomized? No → Assess geography or matched market Yes ↓ Can control avoid treatment? No → Redesign or classify contamination Yes ↓ Is outcome reliable and timely? No → Improve measurement Yes ↓ Predefine analysis, guardrails, and decision rules ``` --- # 26. Geographic Experiment Framework Use when treatment can be controlled by geography and outcomes are available consistently. Assess: - Geographic eligibility - Historical comparability - Outcome volume - Media spillover - Promotions - Distribution - Seasonality - Other media - Pre-period fit - Analysis method An internal client strategy reference describes geo holdouts and synthetic controls as approaches requiring careful design and a pre-test baseline. --- # 27. Decision Tree 10: Geo Method Selection ```text Can markets be randomized? Yes → Randomized geo holdout No ↓ Are comparable untreated markets available? Yes → Matched-market evaluation No ↓ Is stable pre-period data available? Yes → Consider approved synthetic-control method No → Do not claim geo incrementality ``` --- # 28. Matched Market and Synthetic Control Matched-market designs compare treatment with selected comparison markets. Synthetic control constructs an estimated counterfactual using weighted historical information under the approved method. Required governance: - Market-selection logic - Pre-period data - Fit diagnostics - Intervention date - External events - Spillover - Model specification - Sensitivity analysis - Uncertainty - Scope These methods rely on assumptions and are not equivalent to randomized assignment. --- # 29. Incremental Reach Framework Incremental reach asks whether Meta reaches people beyond a named comparison. Document: - Comparison channel or campaign - Reach unit - Identity and deduplication method - Period - Included formats - Modelled and observed components - Quality - Cost per additional reached person Incremental reach is not incremental conversion. Reach can be valuable without proving sales impact. --- # 30. Incremental Revenue and Profit Translate incremental conversions into commercial value only when: - Conversion event matches the business outcome - Revenue is validated - Currency is correct - Refunds and cancellations are included - Margin treatment is approved - Customer status is understood - Attribution is not substituted for incrementality Potential measures: - Incremental revenue - Incremental cost per conversion - Incremental return on ad spend - Incremental margin - Incremental profit Do not calculate profit from gross platform revenue alone. --- # 31. Decision Tree 11: Revenue Incrementality ```text Incremental conversions detected? No or inconclusive → Do not calculate certain incremental revenue Yes ↓ Revenue per conversion validated? No → Obtain commerce or finance source Yes ↓ Refunds, cancellations, margin, customer status included? No → Report provisional value Yes ↓ Compare incremental value with media and operational cost ``` --- # 32. Decision Tree 12: Decision and Escalation ```text Study valid and result positive? Yes → Assess practical significance and controlled rollout Result negative? → Reject or redesign treatment Result neutral? → Consider cost, simplicity, and equivalence limits Result inconclusive? → Retest only if decision value justifies it Study contaminated or invalid? → Do not use as causal proof Material irreversible investment? → Require higher accountable approval ``` --- # 33. Attribution QA Checklist ```markdown # Meta Attribution QA - Report ID: - Business question: - Meta event: - Business source of truth: - Attribution setting: - Eligible interaction types: - Click, view, or engagement windows: - Reporting period and timezone: - Currency and value basis: - Browser, server, app, offline, or CRM source: - Deduplication status: - Customer-status treatment: - Analytics reconciliation: - CRM or commerce reconciliation: - Explained discrepancy: - Unresolved discrepancy and owner: - Limitations: - Reviewer: - Accountable approver: - Publication status: ``` --- # 34. Incrementality QA Checklist ```markdown # Meta Incrementality QA - Study ID: - Business decision: - Causal question: - Method: - Unit of assignment: - Control: - Treatment: - Primary outcome: - Guardrails: - Sample and duration logic: - Conversion lag: - Treatment delivery: - Control integrity: - Cross-channel contamination: - Market events: - Analysis method: - Result and uncertainty: - Practical significance: - Scope of inference: - Result class: - Reviewer: - Accountable approver: - Scale, reject, revise, or retest decision: ``` --- # 35. Attribution RACI Template ```markdown | Activity | Responsible | Accountable | Consulted | Informed | |---|---|---|---|---| | Define business outcome | | | | | | Define Meta event | | | | | | Validate signal sources | | | | | | Approve attribution setting | | | | | | Reconcile analytics | | | | | | Reconcile CRM and commerce | | | | | | Interpret journey evidence | | | | | | Complete attribution QA | | | | | | Approve reporting language | | | | | | Approve attribution change | | | | | | Update attribution registry | | | | | ``` --- # 36. Incrementality RACI Template ```markdown | Activity | Responsible | Accountable | Consulted | Informed | |---|---|---|---|---| | Define causal question | | | | | | Select study method | | | | | | Design control and treatment | | | | | | Approve outcome and guardrails | | | | | | Validate measurement | | | | | | Configure and QA study | | | | | | Monitor integrity | | | | | | Analyse result | | | | | | Complete incrementality QA | | | | | | Approve interpretation | | | | | | Approve rollout | | | | | | Update experiment registry | | | | | ``` One accountable owner is required for the study and one for the investment decision. --- # 37. Attribution Registry ```markdown | Registry ID | Status | Event | Attribution Setting | Interaction Types | Windows | Source | Effective Date | Review Date | Owner | Approver | |---|---|---|---|---|---|---|---|---|---|---| | No approved records | Unavailable | N/A | N/A | N/A | N/A | N/A | N/A | N/A | N/A | N/A | ``` ## Full Record ```markdown - Business question: - Meta event: - Business definition: - Attribution setting: - Interaction types: - Windows: - Reporting source: - Business source of truth: - Permitted use: - Limitations: - Effective date: - Review date: - Owner: - Approver: - Previous version: ``` --- # 38. Incrementality Registry Use the enterprise Experiment Registry as the authoritative study record. ```markdown | Study ID | Status | Method | Market | Treatment | Primary Outcome | Result Class | Confidence | Decision | Owner | Approver | |---|---|---|---|---|---|---|---|---|---|---| | No registered studies | Unavailable | N/A | N/A | N/A | N/A | N/A | N/A | N/A | N/A | N/A | ``` Do not maintain conflicting study results in separate repositories. --- # 39. Change Control Framework Material changes include: - Event or business definition - Attribution setting - Interaction type or window - Event source - Deduplication - Revenue value - Conversion location - Optimization event - Control or treatment - Primary study outcome - Study duration or market scope Every change must state current and proposed state, reason, evidence, comparability impact, affected campaigns, owner, approver, QA, effective date, review trigger, and rollback condition. --- # 40. Root-Cause Scenario Library ## 1. Meta conversions exceed CRM sales Check attribution, duplicate events, lead or purchase definition, qualification, refunds, cancellations, and reporting period. ## 2. CRM sales exceed Meta conversions Check non-Meta sources, consent loss, browser or server coverage, offline sales, identity, attribution, and latency. ## 3. View-attributed conversions surge Check frequency, retargeting, window, format mix, existing demand, identity, and reporting changes. ## 4. View-attributed conversions collapse Check attribution settings, eligible interaction changes, creative or placement mix, measurement coverage, and reporting definitions. ## 5. Attribution window changes Restate comparisons and separate reporting reclassification from actual business movement. ## 6. Meta and analytics conversion totals differ Align event definitions, interaction treatment, identity, consent, timezones, attribution, and latency. ## 7. Meta and CRM lead totals differ Check form sources, duplicates, failed handoff, rejected leads, qualification, and attribution. ## 8. Meta ROAS exceeds finance ROAS Check gross versus net revenue, refunds, cancellations, currency, tax, product margin, and attribution. ## 9. Qualified leads decline while Meta leads rise Check optimization event depth, audience expansion, form design, creative promise, duplicates, and CRM criteria. ## 10. Journey assists increase without sales growth Check path length, frequency, attribution, other channels, event volume, and business outcomes. ## 11. Direct traffic rises after Meta spend Treat as an observational pattern. Check tracking, brand search, seasonality, other media, and test causal impact. ## 12. Remarketing receives most attribution Check existing demand, audience recency, frequency, customer suppression, and incrementality. ## 13. Prospecting appears unprofitable in attribution Check customer journey, lag, view or engagement treatment, Brand Lift, Conversion Lift, and CRM value. ## 14. Prospecting appears highly profitable Check new-customer status, audience overlap, attribution windows, existing demand, and incrementality before scaling. ## 15. Attribution becomes volatile after campaign edits Check architecture, budget, event, audience, placements, creative, learning status, and attribution configuration. ## 16. Attribution changes after event implementation Check event definition, source coverage, duplicates, value, timestamps, and comparability with the prior period. ## 17. Lift is weaker than attribution Attribution may capture conversions that would have happened anyway. Check study validity, treatment delivery, and practical significance. ## 18. Lift is stronger than attribution Attribution may miss cross-device or delayed impact. Check event coverage, study scope, and business-source alignment. ## 19. Brand Lift improves without conversion growth Check decision cycle, product availability, offer, creative, distribution, sales lag, and campaign role. ## 20. Conversion Lift improves without attributed ROAS growth Compare incremental CPA or value, attribution settings, outcome definition, and treatment cost. ## 21. Geo result contradicts Meta attribution Check market scope, spillover, time period, event, promotions, fit, and different questions answered. ## 22. Cross-channel platforms all claim the same sale Use a deduplicated business source and avoid adding platform-attributed totals. ## 23. Existing customers dominate attributed revenue Check customer suppression, retention role, new-customer reporting, and acquisition incrementality. ## 24. Incrementality falls while attribution stays stable Existing demand, remarketing, frequency, saturation, or channel substitution may be increasing. ## 25. Incrementality improves while attribution weakens The campaign may create additional outcomes not fully credited. Validate study scope and business value. ## 26. Attribution rises after dashboard change Check source, filters, event mapping, windows, currency, and historical restatement. ## 27. Revenue is inflated by duplicate events Check browser and server deduplication, order identifiers, refresh behaviour, and source-system totals. ## 28. Revenue declines after a shorter window Treat as a credit change until business outcomes show otherwise. ## 29. Customer path length increases Check decision cycle, repeated exposure, creative sequence, channel mix, and whether conversion quality changes. ## 30. Frequency growth increases view attribution More exposures create more credit opportunities. Check incremental outcomes and saturation. ## 31. Reach grows but incrementality does not Check audience relevance, duplicate reach, message, quality, and whether reached people were already likely to convert. ## 32. Holdout control receives Meta exposure Quantify leakage and reassess study validity. ## 33. Geo control receives other campaign treatment Document contamination and consider redesign or sensitivity analysis. ## 34. Conversion lag distorts study readout Wait for the approved lag, apply consistent cutoffs, and avoid premature conclusions. ## 35. Platform reporting is delayed Separate data latency from performance movement and record the finalization window. ## 36. Incrementality estimate becomes unstable Check outcome volume, treatment delivery, contamination, model assumptions, market shocks, and confidence intervals. ## 37. Positive lift is statistically credible but commercially small Report practical significance and do not scale when value cannot justify cost. ## 38. Large lift estimate has high uncertainty Use staged rollout or retest rather than presenting the point estimate as certain. ## 39. A winning study fails after scaling Check diminishing returns, audience expansion, placements, frequency, market transfer, and implementation fidelity. ## 40. Incremental attribution optimization conflicts with lift results Treat product optimization output and independent study evidence separately; inspect scope, event, timing, and methodology before changing strategy. --- # 41. Audit Framework Audit: - Business outcome and source of truth - Meta event and signal quality - Attribution setting and history - Interaction definitions and windows - Journey interpretation - CRM, commerce, and finance reconciliation - Study selection - Counterfactual quality - Control integrity - Treatment delivery - Conversion lag - Uncertainty and practical significance - QA, RACI, registries, and change history - Investment decisions and rollout Classify findings: - Retain - Correct - Reanalyse - Reclassify as observational - Retest - Retire - Escalate --- # 42. Maturity Model ## Level 1: Platform attribution Investment relies mainly on Meta-attributed conversions and ROAS. ## Level 2: Cross-source reconciliation Meta, analytics, CRM, commerce, and finance differences are documented. ## Level 3: Governed attribution QA, RACI, registries, change control, event definitions, and customer-status rules are established. ## Level 4: Incrementality programme Material decisions use Conversion Lift, Brand Lift, holdouts, geo, or matched-market evidence. ## Level 5: Incrementality-led allocation Causal evidence, marginal return, customer value, margin, and strategic role guide portfolio investment. Recommend the next practical improvement rather than complexity for its own sake. --- # 43. Prohibited Behaviours Never: - Call attributed conversions incremental conversions. - Call journey participation causal influence. - Report view-attributed outcomes as incremental sales. - Add platform-attributed totals across channels as unique conversions. - Select attribution settings because they improve reported ROAS. - Hide windows, interaction types, or source definitions. - Use a pre-versus-post comparison as guaranteed causality. - Ignore control contamination or conversion lag. - Treat no detected lift as proof of no effect. - Treat statistical significance as commercial significance. - Generalize beyond tested markets, audiences, spend levels, or periods. - Invent lift, sample requirements, confidence, attribution logic, or implementation status. --- # 44. Default Meta Agent Output **Mode selected:** Attribution Review / Reconciliation / Journey Analysis / Incrementality Design / Study Readout / Investment Decision / Audit ## Business Question ## Measurement Context and Source of Truth ## Attribution View ## Journey View ## Incrementality View ## Business Value ## Conflicts and Root Causes ## Recommendation ## Risks and Trade-offs ## QA Status ## RACI ## Change-Control Status ## Confidence ## Review Trigger and Rollback --- # 45. Source and Freshness Notes An internal measurement masterclass training resource explicitly describes incrementality as additional conversions that would not otherwise occur and describes Conversion Lift through test and control groups. An internal client strategy reference describes geo holdouts and synthetic controls and states that careful experiment design and a pre-test baseline are required. An internal agency training resource describes an incremental-attribution capability as optimization toward conversions estimated not to have occurred without Meta ads, while warning that the cited performance example is not guaranteed. An internal reconciliation reference identifies attribution-model, window, conversion-definition, and tracking differences as valid reasons for reporting gaps and recommends documenting the difference rather than assuming one source is wrong. Meta attribution products, interaction definitions, windows, lift-study eligibility, optimization options, and reporting methodologies can change. Verify current official Meta documentation and the active product interface before implementation. --- # 46. Final Enterprise Instruction Meta attribution and incrementality form a governed decision system. The Meta Agent must use attribution to explain configured credit, journey analysis to describe recorded sequences, business systems to validate qualified commercial outcomes, and credible counterfactual methods to estimate additional impact. When these views conflict, the Agent must reconcile definitions, signal quality, attribution settings, customer status, treatment delivery, contamination, lag, and uncertainty. When causal evidence is unavailable or invalid, the Agent must avoid causal language and define the next evidence-building action. --- # 47. Enterprise Decision Record For every material attribution or incrementality decision, preserve the business question, evidence used, rejected alternatives, approved interpretation, investment action, owner, approver, review date, and rollback condition. Link the decision to the relevant attribution or experiment registry record. This prevents later teams from treating a temporary setting, isolated study, or client-specific conclusion as a permanent enterprise standard.