The attribution you have is not the attribution you need
Every platform reports the conversions it can see and none of the ones it caused. Here is the small set of things that actually settle the argument.
There is a meeting that happens in most companies about once a quarter. Somebody puts up the platform dashboards, the numbers add up to more conversions than the business actually had, and everyone agrees that attribution is hard before moving on to the next slide. Nothing changes, because nobody in the room has the tools to settle it.
The reason last-click and platform-reported attribution both fail is the same, and it is not technical. Both answer the question which touchpoint was nearest the conversion. Nobody has ever wanted the answer to that question. What you want to know is what would have happened if you had not run the thing at all.
Three tools, in order of how much they cost you
You do not need a data science team to get this right. You need three techniques, deployed in the order of what your business can currently afford.
- Geographic holdouts. Turn a channel off in a set of matched regions for four to six weeks. Compare against the regions where it kept running. It is crude, it is unarguable, and it costs you a small amount of revenue in exchange for the truth.
- Conversion lift studies. Where a platform offers a genuine randomised holdout, use it. Where it offers a “lift study” that is really a modelled comparison, do not.
- Media mix modelling. Worth it once you have two to three years of weekly history and more than four meaningful channels. Below that you are fitting noise and calling it insight.
A geo holdout costs you a fortnight of revenue in two regions. Not running one costs you a year of budget in the wrong place.
What a good holdout actually looks like
The most common failure is not statistical, it is organisational: the holdout runs for ten days, somebody gets nervous, and it is switched back on before the effect could possibly have appeared. Agree the duration in writing and agree who is allowed to end it early, which should be nobody.
Match your regions on pre-period revenue rather than on population or on gut feel. Run for at least four weeks, and longer if your consideration cycle is longer than a fortnight. And decide the success threshold before you look, because after you look it stops being a threshold and becomes a negotiation.
The number to report
Once you have incrementality, stop reporting return on ad spend. Report contribution after media, fees and cost of goods, for incremental conversions only. It is a smaller and much less flattering number. It is also the only one that tells you whether to spend more.
Brand is a media multiplier, and you can measure it
The split between brand and performance is a budgeting convenience, not a description of how buying works. What happens when you stop pretending they are separate.
