The Complete Guide to Marketing Attribution for Ecommerce
Why platform-reported ROAS is wrong, how holdout testing works, and how to find true incremental value per channel. The framework behind every attribution decision Parker runs on Cresva.
Chapter 1: The Attribution Crisis
Marketing attribution is broken. Not slightly off - fundamentally, structurally broken. The numbers your platforms report are systematically inflated, and the entire industry allocates billions of dollars based on data that overclaims by 23% on average.
23%
Average Overclaim
Cross-platform
28%
Meta Inflation
Avg across verticals
18%
Google Inflation
Including branded
35%
TikTok Inflation
Highest variance
Here's the core problem: every ad platform is both the seller of advertising AND the measurer of advertising effectiveness. Meta tells you Meta works great. Google tells you Google works great. TikTok tells you TikTok works great. And because they all use different attribution methodologies - view-through windows, click windows, self-attributed conversions - you can add up all the platform-reported conversions and get a number far higher than your actual total conversions.
We call this the “sum problem.” If Meta claims 1,000 conversions, Google claims 800, and TikTok claims 400, that's 2,200 total. But your Shopify shows 1,500 actual orders. Someone is wrong. In reality, everyone is wrong - they're all overcounting, just by different amounts.
Chapter 2: Why Platforms Lie
“Lie” is strong. The platforms aren't deliberately fabricating numbers. They're using attribution methodologies that systematically favor themselves. There are four primary mechanisms:
View-Through Attribution
Meta counts a conversion if someone saw your ad and purchased within the attribution window - even if they never clicked, never engaged, and would have purchased anyway. If someone scrolls past your ad at 2am and buys from a Google search at noon, Meta claims that conversion.
Impact: High - accounts for 40-60% of Meta's overclaim
Multi-Platform Double Counting
A customer sees a Meta ad, clicks a Google ad, and buys. Both platforms claim the full conversion. Neither reports 0.5 conversions. The same sale is counted twice.
Impact: Medium - affects 15-25% of conversions
Organic Cannibalization
Your most loyal customers were going to buy anyway. But they happened to see an ad or click a branded search result on the way to your site. The platform claims that sale as ad-driven.
Impact: High - especially for branded search campaigns
Algorithmic Attribution Windows
Platforms use different attribution windows (1-day, 7-day, 28-day) and default to the most generous. Longer windows capture more coincidental correlations, not causal relationships.
Impact: Medium - inflates by 10-20% depending on window
The iOS 14.5 Factor
Chapter 3: Attribution Models Compared
Before diving into solutions, you need to understand the landscape. There are four main attribution approaches, each with distinct tradeoffs. The industry is shifting from simpler models toward incrementality-based approaches.
Attribution Model Comparison
How it works
100% credit to the last touchpoint before conversion.
Strengths
Simple, easy to implement, no ambiguity
Weaknesses
Ignores discovery channels, heavily biases toward branded search and retargeting
Our Verdict
Massively undervalues awareness and consideration. Will lead you to over-invest in bottom-funnel.
The ideal approach combines methods: use MMM for strategic quarterly allocation, incrementality testing for validating channel effectiveness, and corrected MTA for daily optimization. No single model is sufficient on its own.
What Parker Does
Chapter 4: Overclaim by Platform
Not all platforms overclaim equally. Based on holdout testing across 100+ ecommerce brands on Cresva, here are the typical inflation rates:
Platform Overclaim Calculator
See how much your platform is likely inflating ROAS.
Reported
4.2x
Overclaim
28%
True ROAS
3.0x
Based on cross-brand holdout testing benchmarks. Actual overclaim varies by vertical, audience, and campaign type.
| Platform | Avg Overclaim | Range | Primary Driver | Worst Category |
|---|---|---|---|---|
| Meta Ads | 28% | 15-45% | View-through attribution | Retargeting campaigns |
| Google Ads | 18% | 8-30% | Branded search cannibalization | Brand campaigns |
| TikTok Ads | 35% | 20-55% | View-through + broad attribution | Awareness campaigns |
| Pinterest Ads | 22% | 12-35% | View-through windows | Home & lifestyle |
| Snap Ads | 30% | 18-45% | View attribution defaults | Younger demographics |
Chapter 5: Holdout Testing - The Gold Standard
The most reliable way to measure true attribution is to stop showing ads to a subset of your audience and measure the difference. This is holdout testing - the gold standard of incrementality measurement.
How to Run a Holdout Test
Define your holdout
Select 10-20% of your audience (by geo, cohort, or random split) to receive zero ads from the channel you're testing.
Run for 2-4 weeks
The test needs enough time to capture full purchase cycles. For higher-AOV products, run longer.
Measure the delta
Compare conversion rates between the exposed group and holdout group. The difference is your true incremental lift.
Calculate true ROAS
Incremental revenue (exposed - holdout) ÷ ad spend = true incremental ROAS. This is always lower than platform-reported.
Apply correction factor
Platform-reported ROAS ÷ true ROAS = your correction factor. Apply this to all future platform data.
Budget Consideration
Chapter 6: Building Your Attribution Model
You don't need a data science team to build a reliable attribution model. Here's the practical framework:
Step 1: Baseline
Run holdout tests on your top 2-3 channels to establish correction factors. Start with your biggest spend channels - the overclaim there costs the most money.
Step 2: Correct
Apply correction factors to all platform-reported data. If Meta overclaims by 28%, multiply all Meta ROAS by 0.72 to get corrected numbers.
Step 3: Unify
Create a single source of truth combining corrected platform data with Shopify/revenue data. This is your de-biased view.
Step 4: Iterate
Re-run holdout tests quarterly. Overclaim rates change with audience saturation, creative mix, and platform algorithm updates.
The key insight: you don't need perfect attribution. You need attribution that's directionally correct enough to make better allocation decisions. Even a rough correction factor - knowing Meta overclaims by “roughly 25-30%” - dramatically improves your budget decisions compared to trusting raw platform numbers.
Chapter 7: Platform-Specific Correction Factors
Based on holdout testing across 100+ ecommerce brands, here are the correction factors by platform and campaign type. Apply these to platform-reported ROAS to estimate true incremental ROAS.
| Platform | Campaign Type | Correction Factor | Example |
|---|---|---|---|
| Meta | Prospecting (broad) | 0.68-0.78 | Reported 3.5x → True ~2.5x |
| Meta | Retargeting | 0.45-0.60 | Reported 8.0x → True ~4.2x |
| Meta | Advantage+ Shopping | 0.70-0.80 | Reported 4.0x → True ~3.0x |
| Non-brand Search | 0.80-0.90 | Reported 5.0x → True ~4.3x | |
| Branded Search | 0.15-0.35 | Reported 12x → True ~3.0x | |
| Performance Max | 0.65-0.75 | Reported 4.5x → True ~3.2x | |
| TikTok | Spark Ads | 0.60-0.75 | Reported 3.0x → True ~2.0x |
| TikTok | In-Feed Video | 0.55-0.70 | Reported 2.5x → True ~1.6x |
Chapter 8: Automating Attribution with Parker
Everything in this guide is what Parker - Cresva's attribution agent - executes automatically. Parker continuously monitors platform-reported data, applies correction factors derived from ongoing holdout testing, and feeds de-biased numbers to Felix (forecasting) and Sam (strategy).
What Parker Does, Continuously
Monitors platform-reported ROAS across Meta, Google, TikTok, and Pinterest
Applies and updates correction factors based on ongoing holdout calibration
Detects when overclaim rates change (seasonal shifts, algorithm updates, audience saturation)
Feeds corrected numbers to Felix for forecasting and Sam for budget allocation
Alerts you when a channel's true incremental ROAS drops below your target
Generates attribution reports comparing platform-reported vs corrected performance
The result: you never make a budget decision based on inflated platform numbers again. Every ROAS figure you see in Cresva has been corrected for platform overclaim, giving you the true incremental picture.
This entire methodology is what Parker runs 24/7 on your data. No manual holdout tests. No spreadsheet correction factors. No hoping the platforms are honest. Just accurate attribution, continuously updated, feeding every other decision in the system.
