Advanced Attribution. Markov & Shapley

Introduction. Why Attribution Models Matter

Ad platforms (Meta, Google, TikTok) all claim credit for conversions. The result: if you add up the conversions reported by each platform, you often get 2 to 3 times more conversions than you actually had.

Reality is more nuanced. A prospect sees a Meta ad, clicks a Google ad, receives a follow-up email, then purchases after seeing a TikTok ad. Who deserves the credit? Everyone : but not equally.

This is exactly what multi-touch attribution (MTA) does: it distributes conversion credit across all touchpoints that contributed to the purchase journey. Metrikia offers 9 models to answer this question, each with its own logic.

The 9 Models Explained

Rule-Based Models (Deterministic)

These models apply fixed rules to distribute credit. Simple to understand, fast to compute, but sometimes reductive.

1. First Touch

Principle: 100% of credit goes to the first touchpoint.

When to use: To measure the effectiveness of your awareness and acquisition campaigns. If you want to know which source brings the most new prospects into your funnel, this is the model for you.

Limitation: Completely ignores what happens after the first contact. A prospect may be acquired by Meta but converted thanks to Google. First Touch will only see Meta.

2. Last Touch

Principle: 100% of credit goes to the last touchpoint before conversion.

When to use: To measure the effectiveness of your closing campaigns. If you're optimizing for immediate conversion (retargeting, urgent promos), Last Touch shows what triggers the purchase.

Limitation: Ignores all the awareness and nurturing work that preceded the conversion.

3. Linear

Principle: Credit distributed equally across all touchpoints.

When to use: When you consider every interaction equally important. Useful for a balanced view without bias toward the beginning or end of the journey.

Limitation: Doesn't reflect reality, an ad click and a sales meeting don't carry the same weight.

4. Time Decay

Principle: The more recent a touchpoint, the more credit it receives. Decay follows an exponential function.

When to use: For short sales cycles where recency is decisive (e-commerce, flash offers). Actions close to conversion count more than those from 3 weeks ago.

Limitation: Systematically undervalues awareness campaigns (which occur early in the journey).

5. U-Shaped (Position-Based)

Principle: 40% to the first touchpoint, 40% to the last, 20% split among middle touchpoints.

When to use: When you want to value both acquisition (first contact) and conversion (last contact) while acknowledging the role of nurturing in between.

Pro tip: U-Shaped is often the best compromise for media teams starting with multi-touch attribution. It gives a balanced view without the complexity of data-driven models.

6. W-Shaped

Principle: 30% to the first contact, 30% to lead creation, 30% to the last contact, 10% split among the rest.

When to use: When lead qualification is a key moment in your funnel. Typical in B2B where the transition from "visitor" to "qualified lead" represents a significant step.

Limitation: Requires Metrikia to identify the lead creation moment in the journey, make sure your touchpoints are properly tracked.

7. Full Path

Principle: 22.5% to the first contact, 22.5% to lead creation, 22.5% to opportunity creation, 22.5% to close, 10% split among the rest.

When to use: For complete visibility across the B2B funnel, from discovery to closing. This is the most granular deterministic model.

Limitation: Requires end-to-end tracking. If stages are missing, the model cannot work correctly.

Data-Driven Models (Probabilistic)

These models use your actual data to calculate each channel's contribution. More complex, but closer to reality.

8. Markov Chain

Principle: Models touchpoints as states in a Markov chain. Each state has transition probabilities to other states (including conversion and drop-off).

How it works:

1. Metrikia collects all conversion paths (and non-conversion paths) from your prospects
2. A transition matrix is built: for each channel, what is the probability of moving to the next channel?
3. Steady-state probabilities are calculated
4. The removal effect is measured: if a channel is removed from all paths, how much does the conversion rate drop?
5. Each channel receives credit proportional to its removal effect

When to use: When you have sufficient data (at least 100 conversions with varied paths). The Markov model reveals channels whose removal would have the greatest impact on your conversion rate.

Pro tip: The Markov model is particularly powerful for detecting "invisible" channels, those that never convert on last touch but are essential in the journey. Typically, display or organic social.

9. Shapley Value

Principle: An approach from game theory. Each channel is a "player", conversion is the "game". The Shapley value calculates the marginal contribution of each player across all possible coalitions.

How it works:

1. For N channels, there are 2^N possible coalitions (subsets of channels)
2. For each coalition, the conversion rate is measured with and without each channel
3. The average marginal contribution of each channel yields its Shapley value
4. For small sets (N ≤ 10), Metrikia computes the exact value
5. For larger sets (N > 10), Metrikia uses Monte Carlo sampling for efficient approximation
6. 95% confidence intervals are provided for Monte Carlo calculations

When to use: When you want the most mathematically fair method. The Shapley value is the only model that satisfies all 4 fairness axioms (efficiency, symmetry, null player, additivity).

Pro tip: If Metrikia displays wide confidence intervals on your Shapley results, it means you have too many channels relative to your data volume. Reduce granularity (group similar channels) or wait until you have more conversions.

Credit Distribution. The LRM Algorithm

All models above calculate percentages. But when distributing an actual conversion value (e.g., $1,000) in cents across 3 channels at 33.33% each, you hit the rounding problem.

Metrikia uses the Largest Remainder Method (LRM) via its CreditDistributor:

1. Each channel receives the integer part of its share (33.33% of 100,000 cents = 33,333 cents)
2. The remaining cents (100,000, 99,999 = 1 cent) are distributed to channels with the largest remainders
3. Guaranteed invariant: the sum of credits always equals the conversion value exactly

This is not a trivial technical detail. Without LRM, rounding errors accumulate and distort your reports. The CreditDistributor is used by all 9 attribution models.

Choosing the Right Model

Configuration in Metrikia

Select a Model

1. Go to Attribution > Settings
2. Choose the default model for your dashboard
3. You can switch between models at any time without losing data. Metrikia recalculates on the fly

Compare Models

The Attribution > Comparison screen displays results from multiple models side by side. This is the best way to spot divergences and understand which channel is over- or under-valued depending on the chosen model.

Results by Campaign

In the detail view of a campaign or ad, the Attribution tab shows the contribution calculated by the active model. You immediately see how much credit this campaign receives and how it compares to the platform's own claims.

Pro tip: Systematically compare Markov and Shapley results. If they converge, your attribution is reliable. If they diverge significantly, your data is likely insufficient, stick with a deterministic model until you have more volume.