Attribution Models

What are Attribution Models?

Attribution models determine how conversion credit is assigned across a user’s touchpoints (ads, email, search, social, direct). Different models tell different stories about which channels drive results—and those stories change your budget decisions.

Formulas / Rules (core types):

  • Last-click (non-direct): 100% credit to the last non-direct touch.
  • First-click: 100% credit to the first touch.
  • Linear: Equal credit across all touches.
  • Time-decay: Increasing weight toward later touches.
  • Position-based (U-shaped): Heavier weights to first and last, remainder spread across middle touches.
  • Data-driven (algorithmic): Credit learned from paths (e.g., Markov removal effects or Shapley values).

Key idea: A model is a lens, not the truth. Use multiple lenses and validate with experiments.

Why it matters?

  • Budget allocation: The model you choose moves money across channels.
  • Bias control: Last-click overvalues brand/search & email; underweights prospecting (Meta, TikTok).
  • Testable decisions: A good model aligns with lift tests and incrementality, not just clicks.

KPIQ Perspective

  • User view: ““Email looks amazing in last-click—should I cut Meta?”
  • Technical view: KPIQ compares rule-based attribution models (last/first/linear/time-decay/position-based), highlights model-sensitive channels (often Brand Search & Email ↑, Meta prospecting ↓), runs simple budget reallocation what-ifs (e.g., shift €X to Meta prospecting) and recommends a primary decision model plus a last-click control for sanity checks. It also flags data-quality gaps (UTM hygiene, missing CAPI/dedup, inconsistent lookback windows).

Actionable Insights

  • ✅ Pick a primary decision model (e.g., Data-driven or Position-based) and keep a control model (Last-click) to monitor bias.
  • ✅ Standardize UTMs & naming; use last-non-direct logic to avoid “Direct” stealing credit.
  • ✅ Set consistent lookback windows (e.g., 7-day click / 1-day view) and compare 7 vs 28 days for sensitivity.
  • ✅ Separate Brand vs Non-brand search—brand will dominate last-click and mask prospecting value.
  • ✅ Validate with incrementality (geo holdout, conversion-lift, PSA) and triangulate with MMM for long-term mix.
  • ✅ Track both channel ROAS and blended ROAS (total revenue ÷ total spend) to avoid tunnel vision.

Practical Example

Scenario (last 30 days): 400 purchases. Paths by share:

  • 30%: Meta → Direct → Purchase (120)
  • 25%: Google Search (Non-brand) → Purchase (100)
  • 20%: Meta → Google Search (Brand) → Purchase (80)
  • 15%: Email → Direct → Purchase (60)
  • 10%: Organic → Email → Purchase (40)

Spend: Meta €8,000 · Search Non-brand €6,000 · Search Brand €2,000 · Email €1,000 · Organic €2,000 (content). AOV: €45.

Model Comparison (conversion credit by channel)

  • Last-click (non-direct): Meta 120 · Search NB 100 · Search Brand 80 · Email 100 · Organic 0
  • First-click: Meta 200 · Search NB 100 · Search Brand 0 · Email 60 · Organic 40
  • Linear: Meta 160 · Search NB 100 · Search Brand 40 · Email 80 · Organic 20
  • Time-decay (2-touch ≈ 30/70): Meta 144 · Search NB 100 · Search Brand 56 · Email 88 · Organic 12
  • Data-driven (example distribution): Meta 150 · Search NB 100 · Search Brand 50 · Email 80 · Organic 20

Impact on ROAS (Last-click vs Data-driven)

Revenue = credited conversions × €45

  • Meta: LC €5,400 → ROAS 0.68× · DD €6,750 → ROAS 0.84×
  • Search Non-brand: LC €4,500 → 0.75× · DD €4,500 → 0.75×
  • Search Brand: LC €3,600 → 1.80× · DD €2,250 → 1.13×
  • Email: LC €4,500 → 4.50× · DD €3,600 → 3.60×
  • Organic: LC €0 → — · DD €900 → 0.45× (if you treat content as spend)

Takeaway: Last-click over-rewards bottom-funnel (Brand, Email) and tempts you to cut prospecting. Data-driven shifts more credit to Meta (assist) and reduces Brand/Email inflation—often leading to better long-term growth.

💡 Tip: Keep Email as a retention engine even if its credit drops under data-driven. Shift incremental budget from over-credited Brand Search to mid-funnel prospecting and creative testing.
📖 Click to open the in-depth analysis

Foundations

An attribution model maps touchpoints to contribution weights. Heuristic models (last/first/linear/time-decay/position-based) impose rules; data-driven models learn weights from observed paths (e.g., Shapley values compute marginal contributions across permutations; Markov models measure removal effects by recomputing path conversion rates after deleting a channel).

Key Concepts

  • Lookback window: The time span to credit touches (e.g., 7 vs 28 days).
  • View-through vs click-through: Including views inflates display/social; restrict or separate.
  • Direct handling: Prefer last-non-direct to avoid direct overwriting real touches.
  • Brand vs Non-brand: Split to avoid cannibalization hiding prospecting value.
  • De-duplication: Server-side events (CAPI) + dedup IDs to prevent double counting.

Advanced Methods

  • Shapley: Fair credit via average marginal contribution across all orderings.
  • Markov: Transition probabilities across states; removal effect ranks channels by necessity.
  • Triangulation: Combine MTA (user-level) with MMM (aggregate) and holdout tests for robustness.
  • Hierarchical models: Partial pooling to stabilize sparse segments.

Common Pitfalls

  • Mismatched lookbacks across platforms → incomparable ROAS.
  • Counting “Direct” as last-click winner → under-crediting top/mid funnel.
  • Not separating brand search → inflated search ROI, starved prospecting.
  • Drawing conclusions without lift tests → model may fit noise, not causality.

Further Reading

  • Shaikh, Ji & Wang — Practical Ad Attribution Modeling
  • Dalessandro & Perlich — Markov Models for Multi-Touch Attribution
  • Lipovetsky & Conklin — Shapley Value in Marketing

 

Back to blog

Resources / Further Reading