An attribution model is a rule set (or algorithm) that splits conversion credit across the user journey. In practice, it assigns value to specific touchpoints—ads, emails, referrals, search—so you can see which efforts actually drive conversion. It isn’t ground truth; it’s a decision lens you choose for budget, bidding, and roadmap calls.
Why it matters
Without a model, teams overfund the loudest last click and starve compounding discovery. Choosing a model clarifies which channel influenced the outcome and which attribution window (lookback period) you’ll standardize for reporting.
Common rule-based models
- Last-click attribution: 100% credit to the final touch. Simple, biased to bottom-funnel.
- First-click attribution: 100% to the opener. Useful for top-funnel discovery.
- Linear (a type of multi-touch attribution): Equal credit to each touch; stable but dilutes strong signals.
- Time-decay: Heavier weight the closer the touch is to conversion (exponential decay with a half-life).
- Position-based (U-shape): Heavy weight to first and last; middle touches share the rest.
Data-driven (algorithmic)
Learns weights from your data using statistical modeling. Works best with consistent UTM parameters and reliable identity stitching across touchpoints.
Quick formula
For a path with touches i=1..ni=1..ni=1..n with weights wiw_iwi and conversion value VVV: crediti=wi∑j=1nwj×V\text{credit}_i=\frac{w_i}{\sum_{j=1}^{n}w_j}\times Vcrediti=∑j=1nwjwi×V
Mini example (A → B → C → Conversion, V=1.0V=1.0V=1.0)
| Model | Credit Split |
|---|---|
| Last-click | A:0, B:0, C:1.0 |
| First-click | A:1.0, B:0, C:0 |
| Linear | A:0.33, B:0.33, C:0.33 |
| Time-decay (half-life=1 step) | A:0.125, B:0.375, C:0.5 |
| Position-based (40–20–40) | A:0.4, B:0.2, C:0.4 |
Practical notes
- Match the model to the decision: prospecting → first/position-based; CRO/retargeting → last/time-decay; portfolio → data-driven.
- Lock your UTM schema and naming before modeling; messy inputs equal noisy outputs.
- Keep the same attribution window for time-series comparisons.