How to Choose the Right Attribution Model

Why Model Selection Matters

The attribution model you choose determines which channels get credit—and therefore budget. Choose wrong, and you'll systematically over-invest in some channels while starving others.

Consider this example with the same conversion data:

Same data, wildly different conclusions. Under last-touch, email dominates. Under position-based, paid social leads. Your budget decisions would be completely different.

There's no "correct" answer—each model reflects different assumptions about how marketing works. The goal is choosing the model whose assumptions best match your business reality.

The Three Decision Factors

Model selection comes down to three questions:

1. How Long Is Your Sales Cycle?

Short cycles have fewer touchpoints, so single-touch models lose less information. Long cycles require multi-touch to see the full picture.

2. How Many Conversions Do You Have?

Data-driven models need volume to learn patterns. With limited data, they'll overfit to noise and give worse results than simple rule-based models.

3. What's Your Primary Use Case?

Different questions need different models. There's nothing wrong with using multiple models for different purposes.

Model-by-Model Breakdown

Last-Touch (Last-Click)

How it works: 100% credit to final touchpoint before conversion.

Best for:
- Short sales cycles (< 7 days)
- Single-session conversions
- Conversion rate optimization
- Comparing against platform reporting

Avoid when:
- Allocating budgets across channels
- Journeys span multiple sessions
- You care about awareness channels

Bias: Over-credits closers (email, retargeting, branded search).

First-Touch (First-Click)

How it works: 100% credit to first touchpoint in the journey.

Best for:
- Measuring demand generation
- B2B pipeline sourcing
- Brand awareness campaigns
- Understanding channel introduction roles

Avoid when:
- Optimizing conversion rate
- Short cycles where first = last
- Journeys start with untrackable touches

Bias: Over-credits introducers, ignores nurturing and conversion.

Linear

How it works: Equal credit to all touchpoints.

Journey: Social → Blog → Email → Search → Purchase
Credit:   20%     20%    20%     20%     (conversion)

Best for:
- Neutral baseline (no built-in bias)
- Long consideration journeys
- When you're unsure which model to use
- First step beyond single-touch

Avoid when:
- Journeys have clearly more important touchpoints
- You have enough data for smarter models

Bias: None inherent—but may undervalue particularly important touchpoints.

Time-Decay

How it works: More credit to touchpoints closer to conversion. Credit decays based on a half-life (commonly 7 days).

Journey: Social → Blog → Email → Search → Purchase
         Day 1    Day 7   Day 14   Day 20   Day 21
Credit:   5%      10%     25%      60%     (conversion)

Best for:
- E-commerce and retail
- Short-to-medium sales cycles
- When recent touches genuinely matter more
- Products with urgency or decay (events, seasonal)

Avoid when:
- First-touch genuinely starts the journey (B2B)
- You want to value awareness fairly
- Sales cycles are very long (decay makes early touches invisible)

Bias: Under-credits early-funnel, over-credits late-funnel (but less than last-touch).

Position-Based (U-Shaped)

How it works: 40% to first touch, 40% to last touch, 20% split among middle.

Journey: Social → Blog → Email → Search → Purchase
Credit:   40%      7%     6%      7%      40%

Best for:
- B2B lead generation
- When first and last touches are strategically important
- Balancing awareness and conversion measurement
- Teams where demand gen and performance marketing need shared credit

Avoid when:
- Single-session conversions (first = last)
- Middle touches genuinely matter (content-heavy journeys)

Bias: Under-credits middle-funnel nurturing.

W-Shaped

How it works: 30% to first touch, 30% to lead creation touch, 30% to last touch, 10% split among others.

Best for:
- B2B with distinct lead capture moment
- When marketing → sales handoff matters
- Companies tracking MQLs as a key metric

Avoid when:
- No clear "lead creation" moment
- B2C without distinct funnel stages
- Simpler funnels that don't need the complexity

Bias: Requires defining "lead creation" (can be arbitrary).

Markov Chain (Algorithmic)

How it works: Calculates each channel's "removal effect"—how much would conversions drop if that channel didn't exist?

Best for:
- High-volume businesses (2,000+ conversions/month)
- When you want data to determine credit
- Sophisticated marketing teams with analytics resources

Avoid when:
- Low conversion volume (overfits)
- You need to explain model to non-technical stakeholders
- You want simple, transparent attribution

Bias: None inherent, but sensitive to data quality and touchpoint definition.

Shapley Value (Algorithmic)

How it works: Uses game theory to fairly distribute credit based on marginal contribution of each channel.

Best for:
- Very high volume (5,000+ conversions/month)
- Academic rigor requirement
- When fairness of credit is paramount

Avoid when:
- Low volume (computational complexity for nothing)
- Practical speed matters (O(2ⁿ⁾ complexity)
- Stakeholders won't trust a "game theory" model

Bias: None inherent, but computationally expensive and opaque.

The Model Selection Framework

Here's a practical framework for choosing:

Step 1: Assess Your Data Maturity

Don't over-engineer. A simple model you trust beats a complex model you don't understand.

Step 2: Match Model to Business Type

Step 3: Define Primary vs Secondary Models

Most mature teams use multiple models:

Be explicit about which model you use for what. Document it. Stick to it.

Common Model Selection Mistakes

Mistake 1: Chasing "The Best" Model

There's no universally best model. Teams waste months researching when they should start with linear and iterate.

Fix: Choose a reasonable default (linear), ship it, learn from results, adjust.

Mistake 2: Using Data-Driven Too Early

Data-driven models (Markov, Shapley, ML-based) need volume. Under 2,000 conversions/month, they'll overfit and mislead.

Fix: Use rule-based models until you have sufficient data. Test algorithmic models against rule-based baselines.

Mistake 3: Never Validating

All attribution is correlational. Without validation, you don't know if your model reflects reality.

Fix: Run incrementality tests quarterly. Compare attributed lift to measured lift. Calibrate your model.

Mistake 4: Changing Models Frequently

Every model change breaks historical comparison. Teams who switch quarterly can never track progress.

Fix: Commit to a primary model for at least 12 months. Run new models in parallel before switching.

Mistake 5: Using One Model for Everything

Last-touch for budget allocation starves awareness. First-touch for CRO ignores conversion.

Fix: Use the right model for each use case. Document which model answers which question.

Validating Your Model Choice

Attribution models are hypotheses about how marketing works. Validate them:

Compare Model Outputs

Run multiple models on the same data:

Channel A: Last-touch says 5%, Linear says 20%, Position says 25%

If models wildly disagree, investigate why. The truth is probably between them.

Check Against Intuition

Does the model output match business knowledge?

• If brand search gets 50% of multi-touch credit, something's wrong
• If email (to existing customers) gets sourcing credit, data is dirty
• If display gets zero credit despite large spend, model may undervalue views

Run Incrementality Tests

The gold standard. Pause a channel (geo-holdout or randomized) and measure actual impact:

Use calibration factors to adjust your attribution model toward reality.

Summary: The Model Selection Checklist

1. Assess your context:

   • Sales cycle length?
   • Monthly conversion volume?
   • Primary use case?

2. Choose a primary model:

   • Short cycle, low volume → Last-touch or linear
   • Medium cycle, medium volume → Linear or position-based
   • Long cycle, high volume → Position-based, consider data-driven

3. Add secondary models for specific uses:

   • First-touch for demand gen
   • Last-touch for CRO
   • Multi-touch for budgets

4. Validate quarterly:

   • Compare multiple models
   • Check against intuition
   • Run incrementality tests

5. Iterate slowly:

   • Stick with primary model 12+ months
   • Run new models in parallel first
   • Document all changes

Further Reading

On Specific Models:
- First-Touch Attribution — When to credit the introduction
- Last-Touch Attribution — When to credit the close
- Linear Attribution — The neutral baseline

Academic Resources:
- Shapley, L.S. (1953) — Original Shapley value paper
- Abhishek et al. (2012) — Markov chain attribution in marketing

Holly Henderson

Co-Founder, mbuzz

Holly Henderson is Co-Founder of mbuzz. With 10+ years in marketing including roles at Westpac, Avon, and Forebrite, she's obsessed with making measurement actually useful.

Harvard Extension School Forebrite Westpac Avon