The Attribution Window Mismatch: How Standard BI Infrastructure Killed a 4.2x ROAS Channel

The Context

I worked with a veterinary equipment manufacturer selling a $1,200 automated prescription dispensing system. The competitive landscape was brutal—we were 6x the price of manual alternatives, and the average sale cycle ran 45-60 days from initial awareness to PO approval.

The Strategic Constraint:

• Target audience: Veterinary practice managers (not individual pet owners)
• Decision-making unit: 1 buyer controlling 500-2,000 units across multiple clinic locations
• LTV: $47,000 (average customer placed 3 reorders over 18 months)
• Acceptable CAC ceiling: $8,500 (5.5:1 LTV:CAC target)

The Experiment & The Failure

We launched a $45,000 Connected TV campaign via The Trade Desk targeting veterinary practice decision-makers across 12 DMAs. Platform tracking showed:

Week 1 Performance:

• Impressions: 2.3M
• Estimated reach: 14,000 decision-makers
• Attributed revenue (7-day window): $4,200
• Dashboard ROAS: 0.09x

We killed the campaign on Day 9.

We reallocated the remaining $32,000 budget to Meta retargeting, which showed immediate "success":

• Week 1 ROAS: 2.1x
• Add-to-cart rate: 8.7%
• Purchase conversion rate: 0.4%

Meanwhile, the CTV campaign we killed had seeded 340 demo requests that converted over the next 6 weeks—we just weren't measuring it.

The Post-Mortem: Diagnosing the Measurement Failure

The problem wasn't the channel. It was the attribution window/decay mismatch.

What the infrastructure measured:

CTV Spend (Day 1): $45,000 
Revenue (Day 1-7): $4,200 
ROAS: 0.09x

What had actually happened:
CTV Spend (Day 1): $45,000

Revenue attribution across decay curve:

Week 1: $4,200

Week 2: $18,900 (adstock decay ~72%)

Week 3: $31,400 (decay ~58%)

Week 4: $28,700 (decay ~48%)

Week 5: $19,200 (decay ~35%)

Week 6: $12,800 (decay ~22%)

Total attributed revenue: $115,200

Actual ROAS: 2.56x

The Technical Fix: Bayesian MMM with Geo-Holdouts

Six months later, working with a different client in a similar position, I architected the measurement system differently:

1. Geo-Based Incrementality Design

• Split 24 comparable DMAs into treatment (CTV) vs. control (suppressed)
• 60/40 split (political compromise—CFO wouldn't accept 50/50 revenue risk)
• 8-week flight to capture 85% of historical conversion window

2. Custom Attribution Infrastructure Built a Python-based MMM using PyMC3 with:

• Adstock transformation: Geometric decay with estimated half-life of 12 days (derived from historical cohort analysis)
• Saturation curves: Logistic function to model diminishing returns
• Bayesian priors: Informed by industry benchmarks for B2B CTV (typical ROAS range: 1.8x - 4.5x)

Model Structure:

# Simplified representation
adstock_rate = 0.6  # 60% retention week-over-week
saturation_point = 85000  # Spend level where returns diminish

transformed_spend = apply_adstock(raw_spend, decay=adstock_rate)
saturated_spend = apply_saturation(transformed_spend, alpha=saturation_point)
predicted_revenue = baseline + (coefficient * saturated_spend)

3. Reporting Layer Rebuild Replaced Google Analytics' 7-day default with custom BigQuery views that:

• Aggregated conversions across 90-day windows
• Attributed revenue using position-based model (40% first-touch, 40% last-touch, 20% distributed)
• Separated "assisted" vs "last-click" revenue to avoid double-counting retargeting cannibalization

The Results

Geo-Holdout Findings (8 weeks):

• Treatment DMAs: 127 conversions
• Control DMAs: 41 conversions (scaled to equivalent population: 68 expected)
• Incremental lift: 59 conversions attributable to CTV

Economic Analysis:

• Incremental revenue: $247,000 (59 conversions × $4,200 AOV)
• Total CTV spend: $58,000
• True incremental ROAS: 4.2x
• CAC: $983 (well below $8,500 ceiling)

The Counterfactual: If we'd maintained the original strategy (killing CTV, doubling down on retargeting):

• Estimated retargeting spend: $90,000
• Incremental conversions: ~12 (based on incrementality test showing 89% cannibalization)
• Incremental revenue: $50,400
• Actual ROAS: 0.56x
• Opportunity cost: $196,600 in foregone revenue

What I'd Do Differently Today

Infrastructure First:

1. Build attribution windows into the BI layer from day one—default to 90-day lookback for any product >$500 AOV
2. Implement geo-based holdouts as standard practice for any brand awareness channel (CTV, audio, OOH)
3. Use Robyn (Meta's open-source MMM) or Meridian (Google's Bayesian MMM) instead of building custom—faster stakeholder buy-in

KPI Reframing: The right success metric wasn't "ROAS in Week 1." It was:

• Cost per qualified enterprise demo (measured at 45 days): Target <$850
• Pipeline velocity: % of CTV-exposed accounts that entered sales conversations within 60 days

Stakeholder Management: The executive dashboard issue wasn't a training problem—it was an API constraint. Platform APIs don't expose adstock-adjusted metrics, so leadership sees "failure" in Week 1 and pulls budget before the curve materializes.

Solution: Pre-commit to measurement windows in writing. "We will evaluate this channel at 8 weeks, not 8 days. Here's the statistical power calculation showing why."

The Broader Lesson

Most marketing teams aren't failing because they pick the wrong channels. They're failing because their measurement infrastructure makes long-term channels look like short-term failures.

If your BI stack defaults to 7-day attribution, you will systematically:

• Overinvest in retargeting (cannibalization masked as performance)
• Underinvest in brand/awareness (delayed conversions appear as "no ROAS")
• Optimize for harvesting existing demand instead of creating new demand

The fix isn't better media buying. It's better measurement architecture.