Attribution Is Dead; Decisioning Isn't: How to Run Growth on Incrementality

Key Takeaways

1. Attribution measures correlation; incrementality measures causation. Only the latter tells you where to move budget.
2. The "Incrementality Stack" layers always-on monitoring, platform signals, attribution, MMM, and experiments—each serving a distinct purpose.
3. Four experiment types cover most needs: conversion lift/holdouts, geo experiments, auction experiments, and (carefully) observational methods.
4. Translate lift results into incremental CPA and incremental ROAS to make budget decisions portfolio-style.
5. A weekly "Growth Measurement Council" ritual keeps experiments running and decisions flowing without slowing execution.
6. Instrumentation quality is the constraint. Event contracts, UTM hygiene, and CRM reconciliation must be airtight before experiments are valid.

Definitions

Why "Attribution Is Dead" (and What That Actually Means)

Let's be precise: attribution isn't useless—it's misused. The claim "attribution is dead" means:

1. Last-click and MTA cannot reliably inform budget allocation. They measure touchpoint occurrence, not causal contribution. A user who saw your ad and converted might have converted anyway.
2. The data foundation is collapsing. iOS ATT, Chrome's Privacy Sandbox, GDPR/CCPA consent fragmentation, and cross-device identity gaps mean MTA models see less than 40–60% of user journeys in many verticals.
3. Platform-reported conversions are inflated. Walled gardens (Meta, Google, TikTok) have incentives to over-claim credit. Overlap, double-counting, and view-through attribution inflate reported ROAS.
4. Selection bias is baked in. Users who see ads are often already more likely to convert (they searched, visited the site, or match high-intent signals). Attribution credits the ad; the truth is the user was predisposed.

Attribution still has value for diagnostics: understanding channel mix, spotting anomalies, and debugging campaigns. But for decisioning—deciding where to put the next dollar—you need incrementality.

What Replaces It: The Incrementality Stack

No single tool answers every measurement question. Instead, build a layered stack where each layer serves a specific purpose and is used at the appropriate cadence.

┌─────────────────────────────────────────────────────────┐
│           CAUSAL DECISIONING LAYER                      │
│  (Budget allocation, scaling rules, portfolio moves)    │
├─────────────────────────────────────────────────────────┤
│           EXPERIMENTS / HOLDOUTS                        │
│  (Conversion lift, geo tests, PSA tests — ground truth) │
├─────────────────────────────────────────────────────────┤
│           MARKETING MIX MODELING (MMM)                  │
│  (Macro trends, channel-level contribution, long-term)  │
├─────────────────────────────────────────────────────────┤
│           ATTRIBUTION (MTA / Last-Click)                │
│  (Diagnostic, journey visualization, anomaly detection) │
├─────────────────────────────────────────────────────────┤
│           PLATFORM SIGNALS                              │
│  (Meta CAPI, Google Ads conversions — fast, biased)     │
├─────────────────────────────────────────────────────────┤
│           ALWAYS-ON MONITORING                          │
│  (Dashboards, pacing, spend vs. KPI directional checks) │
└─────────────────────────────────────────────────────────┘
            

What Each Layer Does (and When It Lies)

Key insight: Use the bottom layers for speed and the top layers for truth. Never let platform signals override experiment results.

The 4 Experiment Types (Deep Dive)

Table 1: Incrementality Methods Comparison

Operating Model: Running Incrementality Weekly Without Slowing Growth

Experiments are useless if results sit in a slide deck. Integrate incrementality into your weekly operating rhythm.

The Growth Measurement Council (Weekly Ritual)

Purpose: Review experiment results, make budget decisions, prioritize next tests.

Attendees: Head of Growth, Performance Marketing Lead, Analytics/BI Lead, Finance partner (optional), Product Analytics (if applicable).

Cadence: Weekly, 45–60 minutes.

Agenda:

1. Dashboards & Anomalies (10 min): Review pacing, spend, conversion trends. Flag any instrumentation issues.
2. Experiment Readouts (20 min): Present completed or in-flight test results with confidence intervals. No tests? Review observational signals.
3. Decision Checkpoint (15 min): Based on results, decide: scale, pause, reallocate, or extend tests.
4. Next Tests Prioritization (10 min): Confirm upcoming experiments, owners, and timelines.

Outputs:

• Documented budget moves (what changed, why, expected impact)
• Updated experiment roadmap
• Action items for instrumentation or data fixes

RACI for Incrementality Operations

Decision Rules: Translating Lift into Budget Moves

From Lift to Incremental Metrics

Incremental CPA = Total Spend / Incremental Conversions

If a campaign spent $100,000 and the lift test showed 500 incremental conversions (above control), incremental CPA = $200.

Incremental ROAS = Incremental Revenue / Total Spend

If those 500 conversions generated $250,000 in incremental revenue, incremental ROAS = 2.5x.

Compare these to your platform-reported CPA/ROAS. The gap is your "attribution inflation" discount.

Scaling Rules (Decision Table)

Portfolio Thinking

Treat channels like a portfolio. Rebalance based on incremental efficiency:

• Rank channels by incremental CPA or ROAS.
• Shift marginal dollars from low-lift to high-lift channels.
• Set confidence thresholds: only move budget when CIs don't overlap.

Tooling + Instrumentation Checklist

Experiments are only as valid as your data. Before running any lift test, verify:

Event Contract for Conversions

• ☐ Conversion events are defined consistently across platforms and CRM.
• ☐ Event firing is validated (no duplication, no missing events).
• ☐ Attribution windows are aligned (7-day click, 1-day view, etc.).

UTM Hygiene

• ☐ UTM taxonomy is standardized and enforced.
• ☐ Auto-tagging is enabled where available; manual UTMs are audited.
• ☐ UTM parameters persist through redirects and form submissions.

CRM Outcome Mapping (Lead → Sale)

• ☐ Lead IDs are passed to CRM with source/medium/campaign.
• ☐ Downstream outcomes (SQLs, opportunities, closed-won) are mapped back.
• ☐ Conversion latency is understood and accounted for in test design.

Deduping and Latency

• ☐ Duplicate conversions are deduplicated across platforms.
• ☐ Conversion windows capture delayed conversions (especially for high-ticket/B2B).
• ☐ Time-zone alignment is verified.

Reconciliation Between Analytics and CRM

• ☐ Weekly reconciliation between analytics conversions and CRM records.
• ☐ Discrepancies are logged and root-caused.
• ☐ Tolerance thresholds are defined (<5% variance acceptable).

Marketing Observability

• ☐ Event volume monitoring with alerts for drops/spikes.
• ☐ UTM parameter coverage tracked.
• ☐ Consent rates and data availability monitored.
• ☐ "Tracking break" incidents are logged and resolved before test analysis.

Common Failure Modes + How to Avoid Them

1. Running too short: Conversion lag means you miss 20–40% of conversions. Run at least 2–4 weeks for user-level, 4–8 weeks for geo.
2. Holdout contamination: Users in holdout see ads via other channels, devices, or retargeting. Audit exclusion logic carefully.
3. Insufficient power: Too few users or geos → wide CIs → inconclusive results. Calculate sample size upfront.
4. Platform-run tests with no transparency: Black-box tests from platforms may have design flaws. Request methodology details; validate with independent tests.
5. Over-relying on observational estimates: DiD and propensity matching are directional, not causal. Never scale budget based solely on observational analysis.
6. Ignoring external factors: Seasonality, competitor activity, or macro shocks can confound results. Use pre/post comparisons and control groups.
7. Analysis paralysis: Waiting for "perfect" data or running too many tests without acting. Prioritize; make decisions; iterate.
8. Treating all channels the same: Some channels (search, branded) have low incrementality by nature (users would have converted anyway). Calibrate expectations.

Examples

Table 2: 90-Day Implementation Checklist

FAQ: Incrementality and Causal Measurement