How to measure marketing after cookies: MMM, incrementality and the end of ROAS as king

What actually broke

First, let's be clear that cookies didn't die overnight. They have been crumbling for a while. Safari and Firefox have blocked third-party cookies for years, Apple's iOS clamped down hard on in-app tracking, and Europe's consent banner means you aren't even allowed to measure a big chunk of your visitors. Google delayed and reversed its plan to switch cookies off in Chrome several times, but that doesn't change the trend: the signal is decaying, and basing a whole company's measurement on someone else's cookies is a growing gamble.

Worse than the technical problem, though, is the mental one. For ten years we ran marketing on last-click attribution because it was to hand and looked precise. But precise it never was. It was just confident.

Why ROAS lies

Let me say it bluntly, because most companies hate hearing it: ROAS is not return, ROAS is credit. It measures the revenue the system assigned to a campaign, not the revenue the campaign actually added. And that is a chasm of a difference. When someone googles your name because they saw your billboard, then clicks your brand search ad, attribution hands all the credit to that last click. The billboard that sent them there gets nothing. Brand search and retargeting thus look like stars, while mostly they are collecting demand that already existed.

The question you actually want answered is not "how much revenue did the campaign get credited with" but "how many sales would not have happened if we switched the campaign off". That is incrementality, lift. And last-click attribution cannot answer it, no matter how you twist it. The economist Grace Kite has said it for years: only modelling over a longer period reveals the true contribution that the click systematically overrates at the bottom and underrates at the top.

Leg one: marketing mix modelling

Marketing mix modelling (MMM) is nothing new, big FMCG companies used it long before the internet. It is a statistical model that estimates from historical data how individual channels, and things outside marketing, contribute to sales: price, season, competition, weather, distribution. Its key advantage today: it works with aggregate data, it does not need to track individuals, so it is immune to the end of cookies and friendly to GDPR.

And here is what is new: MMM has stopped being the preserve of corporations with a million-dollar analytics budget. Google released an open-source tool, Meridian, Meta has open-source Robyn, and firms like Analytic Partners do it as a service. Even a mid-size online retailer can now build a reasonable model that tells it where to put the next crown. MMM is great for strategic questions: how to split budget across channels, how much to brand versus performance, where the diminishing returns kick in. It is not great for fast day-to-day tuning, other tools do that.

Leg two: incrementality and experiments

If you take one word away from this whole piece, make it this one. Incrementality is not estimated from correlations, it is tested by experiment, and an experiment is the closest thing to causality we have in marketing. Two basic flavours. A geo-test: run the campaign in one region and leave it off in another comparable one, and the difference in sales is the lift. A holdout (ghost ads): deliberately withhold the ad from part of the audience and compare how they bought against those who saw it.

The results tend to be sobering, even painful. More than once a channel with a beautiful ROAS turns out to add almost nothing, because it was reaching people who would have bought anyway. That is not a pleasant finding, but it is a finding worth paying for. Better an uncomfortable truth today than a pretty number you burn budget against for a year.

Leg three: first-party data and consent

The third pillar is to stop borrowing other people's data and start collecting your own properly. First-party data is what people leave directly with you: purchases, sign-ups, on-site behaviour, the email relationship. Technically that means server-side tracking, a conversions API and an honestly configured consent mode, so the data is consented and compliant with GDPR. It is not only about compliance, quality first-party data is a competitive advantage today, because rivals cannot copy it from public sources.

Attribution from this data still has a place, but in a different role than the one we built it into for years. Not as a judge that hands out credit and sets budgets, but as an operational compass for tuning campaigns day to day. For big decisions it is short-sighted, for small ones it is handy.

Triangulation: why you need all three

Here is the point WARC, Google, Meta and Analytic Partners now agree on: no single method is the truth. Each sees a different part of reality and has a different blind spot. The answer is called triangulation, three methods side by side, each doing what it is best at:

In practice: MMM tells you how much to put into TV, social and search. Incrementality checks whether the model is lying and calibrates it. Attribution then runs the daily optimisation inside channels. Three tools, three jobs, one truth assembled from different angles.

What to actually deploy in a mid-size company

• 1. A first-party foundation. Server-side measurement, conversions API, consent mode. Without clean data of your own you can't build anything else.
• 2. Lightweight MMM. You don't have to buy an enterprise tool. Meridian or Robyn and a sensible analyst are enough for a first useful model.
• 3. One geo-test a quarter. Pick a channel you doubt and measure its lift. One good test beats a hundred dashboards.
• 4. Brand tracking. Track awareness, associations and preference too. Brand is the part the click will never see.
• 5. New KPIs. Incremental profit and the cost of a genuinely new customer, instead of blind ROAS.

The most common measurement mistakes

• Running the business on ROAS. You optimise for credit, not lift, and keep shovelling money into the bottom funnel.
• No holdouts. Without a control group you never know what would have happened anyway.
• Brand search as the hero. It collects demand something else created and takes the credit for it.
• Measuring only what's easy. Goodhart's law: when a metric becomes a target, it stops being a good metric.
• Ignoring the long effect. Brand advertising pays back over months, attribution writes it off in days.

Frequently asked questions

Sources: WARC (effectiveness and incrementality) · Analytic Partners, ROI Genome · Google, open-source MMM Meridian · Meta, open-source MMM Robyn · Grace Kite (magic numbers / econometrics) · Les Binet & Peter Field, IPA · Gartner (marketing measurement) · IAB Europe. Data-collection legal framework in the Czech Republic: GDPR and guidance from the Czech DPA.

This is what we do: RevOps & CRM and marketing strategy. How to measure brand and performance together sits in our piece on the Multiplier Effect.

Not sure whether your numbers measure lift or just credit? Let's take 30 minutes. We'll look at what your dashboards say and what they leave out.