Attribution model

Long definition

Most conversions involve more than one touchpoint. A user discovers a brand through organic search, returns via a remarketing ad, then converts on a direct visit. Crediting that conversion to a single channel requires a rule.

The standard models:

• Last-click — 100% credit to the final touchpoint. Simple, biased toward bottom-funnel channels (paid brand, direct). Now deprecated in GA4.
• Last non-direct click — 100% credit to the final non-direct touchpoint. The default in Universal Analytics, retired in GA4 in May 2023.
• First-click — 100% credit to the discovery channel. Biased toward top-funnel channels (organic non-brand, paid display).
• Linear — equal credit across all touchpoints. Simple, no source-bias built in.
• Time-decay — exponentially weights recent touchpoints higher. Default half-life is 7 days.
• Position-based (U-shaped) — 40% to first, 40% to last, 20% split among middle. Compromise between first-click and last-click.
• Data-driven — machine-learning model that compares converting and non-converting paths to assign incremental credit. GA4's default since May 2023; uses Shapley-value-style logic on user-level paths.

Attribution model choice changes channel performance numbers significantly. SEO usually gains under data-driven and first-click; paid brand and direct usually lose. A campaign that "doesn't convert" under last-click can carry 30-50% of the credit under data-driven if it consistently appeared in early-stage paths.

The lookback window matters as much as the model. GA4 uses 90 days for paid and organic acquisition by default, 30 days for the touch-channel of the conversion. Conversions outside the window are uncredited regardless of model. The GA4 attribution overview documents the default behavior and how to switch models in reports.

Common misconceptions

• "Data-driven is unbiased." It minimizes a specific bias (single-touchpoint over-credit) but introduces others. It needs a minimum conversion volume to train, and it changes its own weights over time, making month-over-month comparisons unstable.
• "You should pick one model and stick with it." Different decisions need different models. Use data-driven for budget allocation across channels, first-click to value top-funnel content, last-click to debug conversion-step UX. Treat models as lenses, not as truth.
• "SEO performance changes when the model changes." SEO traffic and conversions don't change. Only the share-of-credit changes. A model switch that moves SEO from 15% to 35% of credit is reporting math, not actual performance.
• "Attribution shows you cause and effect." Models distribute credit; they do not prove causation. Incrementality testing (geo holdouts, conversion lift studies) is the rigorous answer to "did this channel drive the conversion or was it riding free?"