marketing-attribution

name	marketing-attribution
description	Deterministic attribution toolkit — MMM, MTA, incrementality testing, method-selection matrix, formulas, decision tree, common pitfalls.
user-invocable	true
allowed-tools	["Read","Glob","Grep","Write"]

Marketing Attribution Skill

Used by analytics-experimentation and media-buyer-bidder to quantify channel contribution to outcome. Three families: MMM (top-down, channel-level, long-horizon), MTA (bottom-up, touchpoint-level, short-horizon), Incrementality (causal, gold-standard, narrow).

Apply the Hardcoding Directive: every attribution claim in a DecisionRecord must cite method, time-window, data window, and confidence interval. No "the campaign drove $X" without method declared.

1. Marketing Mix Modeling (MMM)

Purpose

Estimate channel-level contribution to a KPI (revenue, leads, brand search) over a multi-week / multi-quarter window. Robust to cookie deprecation, signal loss, walled-garden opacity.

Model Structure

KPI_t = baseline_t
      + Sigma_c [ beta_c * Adstock(saturate(spend_c,t)) ]
      + Sigma_x [ gamma_x * control_x,t ]
      + epsilon_t

Components:

Term	Meaning	Typical functional form
baseline_t	Trend + seasonality	Fourier seasonality + trend
Adstock(x)	Carryover / decay	Geometric: x_t + lambda * Adstock_{t-1}, lambda in [0,1]
saturate(x)	Diminishing returns	Hill: x^alpha / (k^alpha + x^alpha), or log(1+x)
beta_c	Channel coefficient	Estimated, must be >=0 (constrain)
control_x	Exogenous (price, promo, weather, holiday)	Linear

Estimation Choices

Method	When
OLS with positivity constraints	Small data, fast prototyping
Ridge regression	Multicollinear channels (TV + OOH + YouTube)
Bayesian (PyMC-Marketing, Meridian, Robyn)	When priors are informative and uncertainty matters for budget decisions

Bayesian is preferred for any budget decision >= $1M. Priors should come from prior MMM runs, meta-analysis, or experiment-calibrated benchmarks.

Outputs MMM Must Produce

1. Channel decomposition: %-of-KPI from each channel + baseline
2. Saturation curves per channel (current spend vs saturation point)
3. mROI per channel at current spend (marginal $ in -> $ out)
4. Optimal reallocation given budget envelope
5. Confidence intervals (Bayesian credible intervals at 80% and 95%)

MMM Pitfalls

• Insufficient variation in spend (channels at flat budgets are non-identifiable)
• Confounding by always-on baseline (must control for organic trend)
• Aggregation bias (national MMM blind to local mix)
• Stale priors (calibrate with recent geo-experiments)

2. Multi-Touch Attribution (MTA)

Purpose

Allocate credit across digital touchpoints in a user's path-to-conversion. Operates on user-level event log.

Rule-Based Models

Model	Credit allocation	When sensible
First-touch	100% to first touch	Brand discovery emphasis
Last-touch	100% to last touch	Closing-channel emphasis (default in many ad platforms)
Linear	Equal split across N touches	Naive, but unbiased between channels
Time-decay	Weight = exp(-(T - t) / half_life)	Late-funnel-weighted
Position-based (U-shape)	40% first + 40% last + 20% middle	Funnel-aware compromise

Data-Driven MTA

• Shapley value: cooperative game theory — credit_c = average marginal contribution of channel c across all subsets of channels. Theoretically correct, expensive (2^N coalitions; approximation via Monte Carlo sampling).
• Markov chain: model journeys as states; channel credit = removal effect (delta in conversion rate when channel is removed from graph).
• Logistic / survival models: estimate per-touch hazard rate.

MTA Pitfalls

• Cookie / device fragmentation undercounts cross-device journeys
• Walled-garden self-reported conversions inflate own channel
• Selection bias: MTA only sees converters' paths
• Last non-direct override masks true direct-traffic value
• Cannot attribute offline / brand-driven lift (use MMM)

3. Incrementality Testing (Causal Gold Standard)

Geo Holdouts (Geo-Lift)

• Randomize geographies (DMAs, ZIPs, postal codes) into test/control
• Run treatment in test markets; hold control dark
• Measure delta-KPI between matched test/control using synthetic-control or DiD
• Requires ~20-30 markets per arm for reasonable power
• Best for: TV, OOH, audio, broad social

Ghost Ads / Public Service Ad (PSA) Holdouts

• Within paid social/programmatic, eligible users randomly assigned to control see PSA instead of ad
• Compare conversion rates between treatment and control
• Requires platform support (Meta Conversion Lift, Google Ads Conversion Lift, DV360)
• Best for: digital paid channels at scale

Conversion Lift Tests

• Same as ghost ads but reported at the campaign level
• Power calculation: need MDE >= 5-10% typically; campaign duration 2-4 weeks minimum

Incrementality Pitfalls

• Spillover (ads in test market reach control-market users)
• Novelty/Hawthorne effects in short tests
• Test-market selection bias (don't cherry-pick high-performing geos)
• Stopping rules: pre-register; never peek

4. Method-Selection Matrix

Rows: channel type. Columns: data-availability tier.

Channel	Rich first-party data	Limited 1P, walled-garden	Offline / no tracking
Paid social	MTA + Ghost-Ad incrementality	Geo holdout	MMM only
Paid search	MTA (data-driven) + brand-search lift test	MTA rule-based	MMM
Display / programmatic	MTA + Ghost-Ad	Geo holdout	MMM
Email / SMS	MTA (direct attribution)	Hold-out cohorts	n/a
Influencer	UTM + promo-code + matched-market	Promo-code only	MMM
TV / OOH	Geo lift + MMM	Geo lift	MMM
Audio (podcast, radio)	Promo-code + geo lift	Promo-code	MMM
Organic / SEO	Last-non-direct MTA + brand search trend	MTA	n/a
PR / earned	MMM + brand search lift	MMM	MMM

5. Decision Tree

Q1: Is the budget decision worth >= $250k or >= 5% of total media?
  YES -> Q2
  NO  -> Use platform-default MTA (last-touch / data-driven)
Q2: Is the channel digital with platform measurement?
  YES -> Q3
  NO  -> MMM (with geo-lift calibration if budget allows)
Q3: Can we run a holdout >= 2 weeks?
  YES -> Run incrementality test, calibrate MTA with result
  NO  -> Data-driven MTA + sensitivity check vs MMM

6. Reporting Standards

Every attribution claim in a MarketBliss artifact must include:

• Method tag: mmm | mta_rule | mta_dd | incrementality_geo | incrementality_ghost
• Window: e.g. "2025-W30 to 2025-W42"
• Confidence: e.g. "80% credible interval +/-12%"
• Calibration source: e.g. "calibrated by Q2 ghost-ad test (lift = 8.4%)"

7. Common Cross-Method Pitfalls

• Attribution gaming: changing attribution mid-campaign to favor a channel
• Double-counting: summing MTA channel credits + MMM channel contribution
• Selection bias: only modeling converters in MTA
• Confounding: not controlling for price, promo, weather, holiday in MMM
• Stale models: MMM trained on pre-iOS14 data applied to current paid social
• Cherry-picking windows: re-running analysis with different start dates

References

• audience-segmentation (cohort definitions feed MTA)
• media-mix-modeling (downstream consumer of MMM output)
• experimentation-design (incrementality test design specifics)