9 min read · 1,817 words
Multi touch attribution cannot be repaired, because the problem was never the tracking. It assigns marketing credit by correlation and has no counterfactual, so it cannot say what would have happened anyway. The alternatives that work are three separate measurements, each answering a different question at a different frequency.
I learned this from a dashboard rather than a book. Some years ago I inherited a reporting setup in an industrial business where every conversion was credited to the last click, which in practice meant a branded search term. On that dashboard, brand search was the best performing channel by a distance and every upper funnel line item looked like waste. A trade campaign that ran for two quarters showed almost nothing. We came close to cutting it. The only reason we did not was that a regional sales head pointed out that the enquiries he was getting had started quoting language from that campaign back to him. The dashboard had no field for that.
What I had was a machine for rewarding the last thing a buyer touched before doing what they were already going to do.
Why did multi touch attribution stop working?
Three things broke it, and only one of them is technical.
The technical one is the tracking substrate. The industry spent five years preparing for the end of third party cookies and then, on 22 April 2025, Google announced it would not deprecate them in Chrome after all. That sounds like a reprieve. It is not. Consent requirements, browser level protections, walled gardens that report on themselves, and buyers who research on a phone and buy on a laptop had already made the cross platform user graph unreliable. Cookies surviving does not reassemble a journey that was never fully observable.
The second is structural. The platforms that hold the data do not share it. Every walled garden marks its own homework, and the sum of self reported conversions across platforms routinely exceeds the total number of actual sales.
The third is the one nobody wants to say out loud. Multi touch attribution was statistically indefensible before any of this. It observes which touchpoints preceded a conversion and divides credit among them by a rule someone chose. There is no control group anywhere in that process. Correlation with a purchase is not evidence of causing it, and a model that cannot distinguish the two is not a measurement, it is a bookkeeping convention.
The consequences show up in the confidence numbers. Forrester found that 64% of B2B marketing leaders do not trust their own organisation’s marketing measurement enough to make decisions with it, and separately that only 6% of B2B organisations qualify as advanced insight driven businesses. Most marketing teams are running dashboards they privately do not believe.
The three question stack
The replacement is not one better model. It is three measurements that answer three different questions, and the mistake most teams make is expecting any one of them to answer all three.
Layer one, allocation. Marketing mix modelling. It answers where the money should go across channels over a year. It is observational, it works at an aggregate level, and it needs two to three years of history to say anything useful.
Layer two, causality. Incrementality testing. Geographic holdouts and randomised experiments. It answers whether a specific spend caused a specific outcome, because it has an actual control group. It is the only layer that produces a genuine causal number.
Layer three, memory. Brand tracking and share of search. It answers whether the brand is getting easier to sell, which the other two layers cannot see because the effect lands outside their observation window. Les Binet’s work presented at the IPA’s EffWorks Global showed share of search correlating with and leading market share across categories, which makes it the cheapest long term signal available.
| Layer | Question it answers | Typical cost | Refresh | What it cannot tell you |
|---|---|---|---|---|
| Mix modelling | Where should next year’s budget go | Free software, paid analyst time | Quarterly | Whether any single campaign caused anything |
| Incrementality test | Did this spend cause this outcome | The cost of the holdout itself | Per campaign | How to split a whole budget |
| Brand tracking, share of search | Is the brand getting easier to sell | Low for share of search, high for a tracker | Monthly to quarterly | Which channel produced the movement |
Run all three and the numbers argue with each other, which is the point. A mix model that disagrees with an incrementality test is telling you the model is misspecified, and you would never have known.
Why should you run an incrementality test before you buy a mix model?
The industry sells these in the wrong order. Vendors lead with mix modelling because it is a licence and a retainer, and incrementality tests are neither.
Run the test first, for three reasons. It is cheaper: a geographic holdout costs you the media you deliberately withhold in the control region, not a software contract. It is faster: a clean read in six to eight weeks against the two to three years of history a mix model needs. And it produces the one number that makes every other model honest, because a modelled coefficient can be calibrated against a measured causal effect but never the reverse.
This is not a fringe view any more. Google’s Meridian, the open source mix model it released publicly on 29 January 2025, is explicitly built to take incrementality experiment results as priors, agnostic of channel. The model is designed on the assumption that you already have experimental evidence. Most teams buying mix models do not.
An incrementality test is the only marketing number you will ever produce that survives an argument with the finance team, because it has a control group and their models do not.
How do you run a geo holdout when the geography is messy?
The textbook version assumes clean, comparable, isolated regions. India does not offer those, and neither do most Asian and Middle Eastern markets, where media spills across state borders, distribution density varies enormously, and the same brand can be a leader in one cluster and unknown two hundred kilometres away.
Four adjustments make it workable. Match regions on outcome history rather than demographics, pairing territories whose sales moved together over the previous eighteen months. Hold out a cluster of smaller markets rather than one large one, so a single distributor problem cannot wreck the read. Run the holdout long enough to cover at least one full purchase cycle for the category, which in capital equipment or steel contracting means months, not weeks. And write down the expected effect size before you start, because a test you interpret afterwards is not a test.
Where true geographic isolation is impossible, switch to a time based design: alternating on and off periods in the same market, with enough cycles to separate the effect from seasonality. It is weaker than a geo split, and it is still infinitely stronger than a last click dashboard.
What does mix modelling cost now that the software is free?
The licence used to be the barrier. It is not any more. Meta’s Robyn is an open sourced marketing mix modelling package from Meta Marketing Science, and Meridian is on GitHub with its methodology open to inspection. Both are free.
What costs money is everything around them: clean spend data reconciled across agencies and platforms, two to three years of weekly history, someone who can specify a model and defend its assumptions, and the organisational patience to accept a result you did not want. In Indian mid market companies the binding constraint is almost always the data, not the modelling. Media spend sits in three agency reports and a finance system that categorises everything as advertising, and the first six weeks of any mix modelling project is spent reconstructing what was actually spent, where, and when.
Budget for that honestly. A model built on spend data nobody has audited produces confident numbers about nothing, which is worse than the dashboard it replaced.
What do you tell a CFO who wants one number?
You tell them the truth, which is that the single number they have been getting was always fictional, and that you are proposing to replace it with a range and an experiment.
That conversation goes better than most marketers expect, because finance teams already work in confidence intervals. They forecast with ranges, they run sensitivities, they discount for uncertainty. The one function that has been bringing them false precision is marketing. A measurement approach that says “this campaign produced between 4% and 9% incremental volume, and here is the control region” is more credible to a CFO than a dashboard claiming a decimal place, not less.
It also gives you the argument for protecting brand spend when budgets tighten. Analytic Partners, working from a database of more than 750 brands across 45 countries, found that brands which cut spend risked losing 15% of their business to competitors who increased theirs, and that brand messaging outperformed performance messaging 80% of the time. That is a defensible reason to hold the line, and it pairs with the case I have made before on what a B2B company should actually spend on brand.
What to do about this in the next quarter
Start with the cheapest thing that produces a real number.
- Pick one campaign with meaningful spend and design a geographic holdout for it. Write the expected effect down first.
- Start recording share of search weekly for your brand and your three closest competitors. It costs nothing and it becomes useful the moment you have twelve months of it.
- Audit your spend data before you buy any modelling. If you cannot produce clean weekly spend by channel for the last two years, that is the project.
- Stop reporting last click as though it means something. Keep it as an operational signal for paid search bidding, which is the one job it is genuinely good at.
- Renegotiate what marketing reports on. The measurement stack only survives if the reporting cycle stops demanding a quarterly causal claim that no honest method can produce.
The reason attribution persisted long after people stopped believing it is that it was the only thing that produced a number fast enough for a monthly review. The fix is partly methodological and mostly organisational, which is the same conclusion I reached about where data driven marketing gets it right and where it fails, and about why awareness is the wrong thing to be tracking in the first place.
My own position is that most companies are one well designed holdout away from knowing more about their marketing than a year of dashboards ever told them, and that the reason they have not run it is that somebody would have to agree to spend less in one region for eight weeks.
Which of your channels would survive a holdout test, if you actually ran one?
-
Previous Post
Personalization Limits: Where Marketing’s Ceiling Is