Add up the conversions Meta claims, the conversions Google claims, and the conversions your email platform claims, and you will usually find you sold your inventory about 1.7 times. Every platform grades its own homework, every platform uses different attribution windows, and every platform is structurally incentivised to claim credit generously.
Multi-touch attribution is the discipline of resolving those competing claims into one honest picture. Done badly, it is an expensive dashboard nobody trusts. Done well, it changes where your next dollar goes. Here is the setup we deploy, layer by layer.
Layer 1: Clean Tracking, or Nothing Else Matters
An attribution model built on broken tracking produces confident nonsense. Before any modelling, three things must be true: every marketing touchpoint carries consistent UTM parameters governed by a written convention; your GA4 property has correctly defined conversions with deduplication; and server-side signals (Meta CAPI, Google enhanced conversions) are flowing so platform data is as complete as privacy rules allow.
Budget two to three weeks for this. Every team wants to skip to the model. The teams that skip rebuild everything three months later.
Layer 2: Choose a Model You Can Defend
The model choice matters less than most vendors claim, with one exception: last-click is disqualifying. It systematically starves discovery channels that start journeys and over-rewards brand search and retargeting that end them.
- Position-based (40/20/40) is a defensible starting point: it credits the touch that opened the journey and the touch that closed it, with the middle shared
- Data-driven models (GA4's default) are stronger with sufficient volume, roughly 3,000+ monthly conversions, but are a black box you cannot fully interrogate
- Time-decay suits short sales cycles; linear suits almost no one but is at least honest about its ignorance
Layer 3: Calibrate Against Ground Truth
Here is the step that separates working attribution from decorative attribution: incrementality testing. A model is a hypothesis about credit. Experiments are evidence. Geo-holdouts (pausing a channel in matched regions), audience splits, and spend step-tests tell you what actually happens to revenue when a channel's spend changes.
Run one test per quarter on your largest or most doubted channel. When the model and the experiment disagree, the experiment wins and the model gets recalibrated. Over a year, this loop turns your attribution from a claimed number into an earned one.
A model is a hypothesis about credit. An experiment is evidence. When they disagree, the experiment wins.
Layer 4: Report Three Numbers, Side by Side
The reporting layer should show platform-claimed conversions, modelled attribution, and blended reality (total revenue over total spend, MER and blended CAC) next to each other. The platform number tells you what the algorithm thinks it did. The model distributes credit. The blended number is the one your CFO already believes.
When all three trend in the same direction, decisions are easy. When they diverge, that divergence is itself the insight: it usually means a channel is claiming credit for demand it did not create. That is your next incrementality test, and often your next budget cut.
What This Buys You
In practice, brands running this setup typically find 20 to 30% of spend is producing far less incremental revenue than platform reporting suggested, most often in retargeting and brand search. Reallocating that spend toward genuinely incremental channels is the single highest-ROI analytics project most marketing teams will ever run.