Marketing Attribution

What attribution actually does

Marketing attribution assigns conversion credit to the marketing interactions that preceded a conversion. When someone clicks a search ad on Monday, sees a retargeting ad on Wednesday, and purchases on Friday, attribution is the set of rules deciding which touchpoint — or which combination — gets credit for that purchase.

The crucial framing: every attribution model is an interpretation, not a measurement. The underlying data is incomplete by construction — cross-device journeys break identity, privacy frameworks suppress tracking, walled gardens don't share user-level data, and offline influence is invisible. A model takes the observable fragments and applies a convention for distributing credit. Different conventions produce different numbers from the same reality, which is why "what's our real ROAS?" has no single answer — only answers relative to a stated model and window.

The model landscape

Rule-based models apply a fixed convention:

• Last click — full credit to the final clicked touchpoint. Simple, widely defaulted, and structurally biased toward bottom-funnel channels like branded search and retargeting.
• First click — full credit to the first touchpoint; the mirror bias, toward discovery channels.
• Linear / time-decay / position-based — credit spread across touchpoints evenly, by recency, or weighted toward first and last. Less arbitrary in feel, equally arbitrary in substance.

Data-driven attribution (DDA) uses the platform's conversion data to estimate each touchpoint's contribution by comparing converting and non-converting paths. Google Analytics 4 defaults to DDA. It is generally better than fixed rules but remains bounded by what the platform can observe — and it is not auditable from outside.

Attribution windows matter as much as models: a 7-day click window and a 28-day click-plus-view window can double-count or drop entire categories of credit. View-through attribution — credit for impressions that were seen but not clicked — is where platform reports inflate most aggressively.

Beyond user-level attribution sit two complementary methods: media mix modeling (MMM), which infers channel contribution statistically from aggregate spend and outcome data, and incrementality testing (geo holdouts, conversion lift studies), which measures what attribution cannot: whether conversions would have happened without the ad.

Why this determines budget quality

Each ad platform reports conversions using its own model, its own window, and credit for its own touchpoints only. Sum the platform numbers and you get more conversions than actually occurred. Compare platforms with different windows and the "winner" is often just the one with the more generous counting. Teams that shift budget based on raw platform dashboards systematically over-fund channels that harvest existing demand and under-fund channels that create it — the ROAS numbers look better precisely where incrementality is lowest.

In app marketing the problem is sharper: an MMP such as AppsFlyer or Adjust exists specifically to be the neutral referee, deduplicating install credit across networks under privacy frameworks like SKAdNetwork that limit user-level visibility.

Setting up attribution you can act on

1. Get conversion tracking right first. Attribution distributes credit among events; broken or duplicated events poison every model equally.
2. Pick one neutral source of truth for cross-channel comparison — GA4 for web, an MMP for apps — and accept that platform dashboards are for in-platform optimization only.
3. Document the model and window in every report. "ROAS 4.2 (GA4 DDA, 30-day click)" is a usable number; "ROAS 4.2" is not.
4. Align windows to your sales cycle. A 7-day window on a 60-day considered purchase silently deletes most of your funnel's credit.
5. Audit view-through credit. Check what share of each platform's reported conversions are view-through, and discount accordingly when comparing.
6. Calibrate with incrementality tests on your biggest line items annually or when stakes are high. Use lift results to set per-channel discount factors on attributed numbers.
7. When you change models, restate history. Otherwise every trend chart shows a fake step-change at the switch date.

Common mistakes

• Treating platform reports as ground truth. Each platform is an interested party counting its own contribution under its own rules.
• Ignoring view-through credit. Impressions can carry real influence, but unexamined view-through is the single largest source of inflated platform numbers.
• Changing models without documenting impact. A model switch reshuffles apparent channel performance; undocumented, it reads as channels suddenly improving or collapsing.
• Equating attributed with incremental. Branded search and retargeting attribute beautifully while often adding little; attribution cannot detect the difference — only holdout testing can.
• Over-engineering before basics work. A multi-touch model on top of untested tracking and undefined conversion events is precision theater.

FAQ

Which attribution model should I use? For cross-channel reporting, data-driven attribution in a neutral tool beats fixed rules. For platform bidding, the platform's own model is usually fine — automated bidding optimizes against it natively. The discipline that matters is consistency and documentation, not model sophistication.

Why don't my platform numbers add up to my analytics numbers? Because every platform claims conversions it touched, journeys touch multiple platforms, and windows differ. Overlap means platform-reported conversions will exceed actual conversions more or less permanently. This is structural, not a bug to fix — it's why a neutral source exists.

Is last-click attribution dead? As a default for budget decisions, it should be — it systematically misprices upper-funnel work. As a stable, simple convention everyone understands, it survives, and for businesses with short single-channel journeys its distortion is small.

How do privacy changes affect attribution? Consent requirements, tracking prevention, and frameworks like SKAdNetwork shrink the observable share of journeys. Platforms fill gaps with modeled conversions — estimates, marked as such in some tools and silent in others. Practical consequence: user-level attribution degrades over time, and aggregate methods (MMM, lift testing) regain importance. The paid acquisition path covers how these layers fit together.

What's the difference between attribution and incrementality? Attribution distributes credit for observed conversions among observed touchpoints. Incrementality asks whether the conversion would have happened without the ad at all. A channel can attribute thousands of conversions while driving few incremental ones — measuring the gap is what holdout experiments are for.