/pulse

Attribution Modeling

Purpose: assign conversion credit across the touchpoints a user encountered before converting. This is observational credit assignment over tracked events — it is distinct from causal incrementality (experiment) and from aggregate spend allocation (MMM). Use this reference to pick an attribution model, understand its bias, and route correctly to the causal toolkit when the question is actually causal.

Contents

• Where attribution sits (boundary map)
• Rules-based models
• Algorithmic models (Shapley, Markov, GA4 DDA)
• Selection guide
• Privacy / consent caveats
• Implementation notes

Where Attribution Sits — Boundary Map

Three different questions, three different tool families. Do not mix them.

Key rule: MTA answers credit, not causation. A channel can win attribution credit while contributing zero incremental conversions (e.g. branded-search capturing already-decided users). When a stakeholder asks "is this channel worth it?", that is an incrementality question — hand off to experiment, do not answer it with MTA alone.

Rules-Based Models

Deterministic heuristics. Cheap, transparent, but encode a fixed prior about where value lives.

Use rules-based only when (a) path data is thin, (b) you need an explainable model for non-technical stakeholders, or (c) as a baseline to compare an algorithmic model against.

Algorithmic Models

Data-derived credit from observed path patterns. Require sufficient path volume (rule of thumb: thousands of converting and non-converting paths) and break down on sparse or heavily consent-truncated data.

Shapley Value

• Cooperative game theory: each channel is a "player"; credit = its average marginal contribution across all channel coalitions.
• Strength: fair, axiomatic (efficiency, symmetry, null-player, additivity); order-independent.
• Weakness: combinatorial cost grows with channel count (2^n coalitions — collapse rare channels into an "other" bucket); treats presence/absence, ignores sequence and frequency.
• Use when channel set matters more than order, and channel count is moderate (≤ ~12).

Markov Chain (removal effect)

• Model paths as a first- (or higher-) order Markov chain between channel states (incl. start, conversion, null). A channel's credit = the removal effect: the drop in conversion probability when that state is removed from the graph.
• Strength: captures sequence and transition structure; handles loops and repeat touches.
• Weakness: first-order forgets long-range order; needs enough transitions per edge; higher-order models explode in state count.
• Use when order/sequence carries signal (e.g. nurture sequences, retargeting loops).

GA4 Data-Driven Attribution (DDA)

• Google Analytics 4's built-in algorithmic model (Shapley-derived, counterfactual-based).
• Became GA4's default after rules-based models (first/linear/time-decay/position) were deprecated in GA4 reporting (2023).
• Strength: zero-build, integrated with Google Ads bidding.
• Weakness: black-box, Google-walled, conversion-volume thresholds before it activates, cross-platform blind. Treat as a managed Shapley variant, not a separate theory.

Selection Guide

Is the real question "did it CAUSE conversions?"  → experiment (incrementality). STOP.
Is it "how to split budget across channels?"      → MMM (experiment + funnel-cohort-analysis). STOP.
Otherwise (credit assignment over paths):
  Path volume thin / need explainability          → rules-based (last-touch baseline)
  Channel SET matters, order doesn't, ≤~12 chans  → Shapley
  SEQUENCE / loops / nurture order matters         → Markov removal-effect
  On Google stack, want zero-build + Ads bidding   → GA4 DDA
Always: validate the algorithmic model against a rules-based baseline before trusting it.

Privacy / Consent Caveats

MTA degrades under the same cookieless conditions documented in privacy-consent.md and funnel-cohort-analysis.md:

• Consent-denied EEA/UK traffic produces truncated paths — touchpoints silently drop, biasing credit toward observable (often last/direct) touches.
• iOS / cross-device journeys fragment paths; algorithmic models trained on fragments inherit the fragmentation bias.
• Always disclose the consent-denial blind spot alongside any attribution report (mirror the disclosure rule in funnel-cohort-analysis.md).

When path data is too truncated to trust MTA, escalate to aggregate methods (MMM) or causal methods (incrementality) rather than reporting biased per-path credit.

Implementation Notes

• Attribution needs clean touchpoint events: every marketing interaction must emit a consistent event with channel, campaign, timestamp, and a stitched user_id / session_id. Gaps here corrupt every model downstream (see event-schema.md).
• Tooling: GA4 DDA (managed); ChannelAttribution (R, Markov); custom Shapley via Python (itertools coalitions for small n); warehouse-native path tables in BigQuery/Snowflake.
• Report attribution as a range across models, not a single number — divergence between last-touch and Shapley/Markov is itself the insight (it quantifies funnel-stage bias).