Marketing Attribution Data-Driven Models

9 min read

Learn

Quick Reading

Estimated 9 mins

Prereq

Intermediate

Basic ML concepts helpful

Interactive

Static Playbook

Static guide & reference tables

From Rule-Based to Data-Driven Attribution

Rule-based attribution models — first-touch, last-touch, linear, time-decay, position-based (U-shaped) — assign credit using a fixed, pre-decided formula applied identically to every conversion path, regardless of what the data actually shows about which touchpoints matter. Data-driven attribution (DDA) instead uses statistical or machine-learning methods to estimate each touchpoint's actual contribution by comparing converting paths against non-converting paths, learning which sequences and channel combinations are genuinely associated with conversion rather than assuming a fixed weighting scheme upfront.

How Google's Data-Driven Attribution Actually Works

Google Ads' DDA (now the default model since rule-based models were deprecated in GA4 and Google Ads) uses a counterfactual approach built on a modified Shapley value framework: it compares the conversion probability of paths that included a given touchpoint against otherwise-similar paths that didn't, across your actual historical conversion data, to estimate each touchpoint's incremental contribution. This requires a meaningful volume of conversion data to train on — Google requires a minimum conversion threshold (historically around 300 clicks and conversions within a lookback window) before DDA can be applied to a given conversion action, falling back to a default model below that threshold.

Attribution model comparison

Model	Logic	Main Weakness
Last-touch	100% credit to final touchpoint before conversion	Ignores every upper-funnel contribution entirely
Linear	Equal credit across every touchpoint	Treats a brand-awareness impression as equal to a direct click
Time-decay	More credit to touchpoints closer to conversion	Still arbitrary — decay rate is a guess, not learned
Data-driven (Shapley-based)	Credit learned from comparing converting vs non-converting paths	Needs volume; still correlational, not a true causal experiment

Data-Driven Attribution Is Still Not Incrementality

Even a sophisticated data-driven model is fundamentally observational — it explains patterns in historical conversion paths, but a touchpoint can show strong statistical association with conversion without causing any incremental conversions (e.g., a retargeting ad shown to someone who was already going to buy). Only a randomized incrementality test (geo holdouts, ghost ads, PSA tests — covered in other lessons) establishes true causal lift. Treat DDA as the best available tactical, continuous signal, and incrementality testing as the periodic causal calibration check against it.

Markov Chain Models as an Alternative Approach

Outside Google's specific implementation, many marketing analytics teams build their own data-driven attribution using Markov chain models: representing the customer journey as transitions between channel-states, then computing each channel's 'removal effect' — how much total conversion probability drops if that channel were removed from all paths entirely. This method is popular because it's transparent and reproducible with your own first-party data (via BigQuery or a data warehouse) rather than relying on a platform's opaque proprietary model, at the cost of needing meaningful in-house data science investment to build and maintain correctly.

conversion_paths.sql

sql

-- Build conversion path sequences per user from a unified
-- event stream, as input to a Markov chain attribution model.
SELECT
  user_pseudo_id,
  ARRAY_AGG(channel ORDER BY event_timestamp) AS path,
  MAX(IF(event_name = 'purchase', 1, 0)) AS converted
FROM `project.analytics_123456789.unified_touchpoints`
WHERE event_timestamp BETWEEN TIMESTAMP('2026-09-01') AND TIMESTAMP('2026-09-30')
GROUP BY user_pseudo_id

Operational Pitfalls: Cross-Device and Cross-Platform Gaps

Any data-driven attribution model is only as good as the identity graph stitching touchpoints together into a single path. If a visitor sees a display ad on mobile, clicks a search ad on desktop, and converts on a third device without being matched as the same person across all three, the model sees three separate, shorter, less informative paths instead of one true journey — understating upper-funnel contribution systematically. This is a bigger problem the more your signal relies on cookies (third-party cookie deprecation only worsens it) versus authenticated, logged-in cross-device identity.

Don't Mix Attribution Windows Across Compared Channels

A common analysis error: comparing a channel using a 30-day click attribution window against another using a 7-day window, because each platform's default differs. Normalize attribution windows across every channel before comparing performance, or the comparison is measuring methodology differences, not channel performance differences.

What's Next?

The Shapley value approach underlying Google's DDA deserves its own deeper look — the next lesson walks through the actual mathematics of Shapley value attribution.

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →