Attribution after cookies: a practical set-up
GA4, server-side tagging and a simple marketing mix model. A measurement set-up a finance team will accept, built in the order that works.
Third-party cookies are mostly gone. Consent banners take out a share of the rest. Every ad platform now models the conversions it cannot see, and each one models generously. The result is that the platforms together claim more sales than the business made, and the finance team stops believing any of it.
This guide sets out a measurement stack that works within those limits. Build it in this order.
Layer one: get the plumbing right
Everything else sits on GA4 and the platform conversion tags. Most accounts have at least one of the following wrong.
- Consent mode. Set up consent mode v2 properly, so Google can model the conversions of users who decline. Without it, a large share of conversions disappear and the bidding gets worse.
- Server-side tagging. Move the GA4 and Google Ads tags to a server container. It reduces the data lost to browsers and ad blockers and gives you control over what is sent where. It costs a few pounds a month to host.
- Enhanced conversions. Send hashed first-party data (email, phone) with the conversion so Google can match conversions it would otherwise miss.
- Reconcile. Every month, compare GA4 purchases, Google Ads conversions and the order system. Write the three numbers down. The gaps tell you what is still broken.
Do not move to layer two until the reconciliation gaps are under ten per cent.
Layer two: attribution you can explain
Platform attribution answers one question: of the sales we can see, which touchpoints were involved? It cannot see everything, and it cannot tell you what would have happened without the ad. But it is useful for day-to-day bidding and for comparing campaigns within a channel.
- Use data-driven attribution in GA4 and Google Ads where the volume supports it. Below a few hundred conversions a month, use a position-based model and say so.
- Write down the model, the conversion window and the counting method. One page. Give it to finance. The point is not that the model is right. It is that everyone knows what it is.
- Stop comparing channels on platform attribution. Google, Meta and the affiliate network will never agree, because they are each counting their own touchpoints. Use platform data to optimise within a channel. Use layer three to decide between channels.
Layer three: incrementality
The question the board asks is “what would we lose if we stopped this?” Only incrementality tests and marketing mix modelling answer it.
Geo tests. Turn a channel off in some regions and keep it on in matched regions. Compare sales. This is the cleanest test available and it needs no cookies. It needs enough regions to match, at least four weeks, and a channel big enough to move sales measurably. Run one per quarter on the biggest channel.
Marketing mix modelling. A statistical model that explains weekly sales using media spend by channel, price, promotions, seasonality and other factors. It measures every channel at once, including TV, radio and the ones with no click. It needs two years of weekly data and a clear-eyed analyst. Open-source tools like Meridian and Robyn make it affordable for businesses spending a few million a year. Joseph built and ran MMM for a betting and gaming group, where it measured a +9.0 ROI from YouTube brand investment over an 18-month cycle and informed the wider performance programme.
Holdouts. For remarketing and paid social, hold back a random share of the audience and compare. Cheap, fast and usually humbling.
Putting it together
- Layer one runs the bidding.
- Layer two runs the weekly optimisation within channels.
- Layer three sets the budget between channels, once a quarter.
Each layer answers a different question. The mistake is asking one layer to answer all three.
The finance conversation
Bring finance in at layer one. Show them the reconciliation. Show them the one-page attribution note. Show them the geo test design before you run it. A finance director who has seen the method will accept the result. One who is only shown the result will not.
Getting help
The measurement and attribution service covers all three layers, from fixing the tags to running the first marketing mix model.
Talk to Joseph about your account
A thirty-minute call. You describe the account, Joseph tells you what he would look at first. No deck, no pitch.