Attribution Model

An attribution model is a rule for assigning credit across all the marketing touchpoints a user saw before they converted. The same conversion can give 100% credit to the last ad clicked, or split credit across every touch, or weight by recency — and each model produces dramatically different reports about which channel “drove” growth.

The common models, briefly

• First-touch — 100% credit to the first interaction. Good for brand-awareness reporting, bad for short-funnel decisions.
• Last-touch — 100% credit to the last interaction before conversion. Default in most ad platforms. Easy to compute, tends to over-credit retargeting and search.
• Linear — equal credit across every touchpoint. Fair but uninformative.
• Position-based (40/20/40) — 40% to first, 40% to last, 20% split across the middle. A compromise.
• Time-decay — recent touchpoints get more credit than older ones. Reasonable for short consideration cycles.
• Data-driven — uses statistical methods (often counterfactual analysis) to estimate causal contribution. The gold standard, the hardest to implement.

The biggest mistake is treating any of these as truth. They’re lenses. The right model depends on what decision you’re trying to make.

Why this matters for the data team

Different teams will pick the model that makes their channel look good. Brand picks first-touch. Performance picks last-touch. Lifecycle picks linear. Without a documented attribution definition, every weekly report is implicitly arguing for a different budget allocation.

Pick one model as the primary, document the SQL, version it. Use the others as secondary lenses to sanity-check.

How Datost handles attribution

Datost reads your touch data (Segment, ad platforms, CRM) and your documented attribution definition. Ask “what’s our last-touch attribution by channel for Q4 conversions?” and you get the table with the SQL attached. The harder question — “what does our data-driven attribution model say for the same period?” — requires the statistical model to already exist; Datost can run the query against it but cannot invent the model from scratch.