What does marketing attribution cover?
What breaks first when cookies go away?
Last-click reporting in the ad platform, then the analytics property, then anything that relied on a browser pixel at checkout. Conversion events have to come from the server that already knows the customer has paid.
How should a conversion event be mapped?
One event, one currency, a value finance recognises, and a clear split between first and repeat purchases. If 'purchase' means sign-up in one account and first payment in another, the reporting is fiction.
What is a conversion API actually for?
It is a channel for sending events the ad platform will accept. Match rate, event quality and the delay window matter more than the vendor's slides. We go into detail in our article on server-side tracking.
How do you avoid two sources of truth?
Conversions reported by the platform, by analytics and by the warehouse will never match to the euro. They must agree in direction and in definition. We publish the join rules so media cannot claim a win nobody else can see.
Where does modelling still help?
In market and channel mix decisions when user-level joins are thin. Models are labelled as models, and they never replace an order-system event.
How is marketing attribution measured?
| KPI | What it tells you | When we act on it |
|---|---|---|
| Conversion match rate | Share of recorded first purchases that reach the ad platforms. | When it falls below the threshold agreed in week one. |
| Event freshness | Delay between the order system and the platform. | When the delay exceeds the optimisation window. |
| Definition drift | Number of competing conversion definitions in use. | At two. The rule is one. |
| Unattributed conversion share | Sales with no usable source. | When the share grows after a platform change. |
What happens in the first 90 days?
- Days 1–15: inventory of pixels, APIs, warehouse, order-system events and affiliate postbacks. We write the single conversion definition.
- Days 16–40: server-side conversion event live, conversion APIs fed, obvious double counting removed.
- Days 41–70: a match-rate dashboard, plus notes on discrepancies that media can act on without a meeting.
- Days 71–90: the definitions are locked. Media and CRM use the same event, and any models are documented as models.
Where does attribution usually go wrong?
A pile of Google Tag Manager tags instead of an order event
Tags on thank-you pages miss in-app purchases, retried payments and customers who pay from another device. The source of truth is the order system, not the dataLayer.
Optimising platforms on a diluted event
If 'purchase' includes failed payments and abandoned trials, Meta, Google, TikTok and Snap will all find you more of both. Event quality sets campaign quality.
Waiting for perfect identity
A join key that works for 70% of first purchases this month beats a six-month customer data platform (CDP) project that ships nothing. We build attribution as a series of reliable pipes, which is why it sits next to paid acquisition rather than in a separate data project.
Related services: paid acquisition that can only scale once the conversion event is real, CRM that needs the same customer identity the ads use and affiliate management, where partner conversions have to reconcile. For the numbers behind this approach, read server-side tracking, with a worked match-rate example.