Decision first
Start from the business decision the measurement has to inform, then work back to the evidence it needs.
Guide
Attribution, experiments or mix models?
They answer different questions, so a good measurement plan uses them together. Attribution shows where conversions happened, experiments show what the media caused, and mix models set the budget across every channel, kept honest by experiment results.
Questions that sound like one
The budget has to come down before the next flight, and something has to give. Paid social looks strongest in its own platform report, so the marketing lead wants to cut somewhere else. The finance lead asks what paid social added that wouldn't have happened anyway, and nobody in the room can say. Both are asking reasonable questions. They are asking different ones.
- Where did conversions happen? That is attribution.
- What did the media cause? That is incrementality, and it takes an experiment.
- Where should the next dollar go? That is a marketing mix model, calibrated with experiments.
What each method needs, and how far it proves cause
Attribution
Needs clicks and conversions tied together, ideally from one deduplicated source. It reads patterns in the path, so on cause it is weak. Use it for daily optimization.
Experiments
Need a control group, a test sized before spend and a clean comparison. On cause they are strong. Use them to confirm what a channel adds and to calibrate the models.
Mix models
Need spend or exposure and outcomes over time, with prices, promotions and seasonality. On cause they are as good as their data and their calibration. Use them for the budget across every channel, digital and traditional.
Attribution, model by model
Last-click gives all the credit to the final click, which rewards whoever closes, such as brand search and retargeting, and ignores whoever created the demand. First-click does the opposite. Rule-based models split credit by a fixed rule, and the rule is an assumption. Data-driven models split credit by estimated contribution, and they are still observational. Results reported by each platform count what that platform can see, so across platforms the total often runs higher than actual sales. How we read those reports is on Digital advertising.
Two habits keep attribution useful. The attribution window is agreed before launch and stated with the results, and view-through conversions are always reported apart from conversions after a click.
Experiments, and why they're sized first
An experiment keeps a comparison group away from the campaign and measures the difference. It can be a holdout group of people or customers, a geo experiment across regions, a pair of matched markets, or, when no test was run, a synthetic control built from other markets or periods. Regional tests suit channels without individual identity, such as TV, radio and out of home.
The effect of advertising is usually small next to the everyday noise in sales. A test too small to see a real effect is budget spent for nothing, so the minimum detectable effect, the size of the control and the duration are set before any money goes out. Incrementality is a habit, since results move with competitors, seasons and creative.
Mix models, and what keeps them honest
A mix model relates outcomes to spend or exposure by channel over time, alongside prices, promotions and seasonality. It needs no individual tracking, which is why it holds up as browser signal thins. It describes how an ad's effect lingers after it runs (adstock), how each extra dollar in a channel returns less (saturation), and so how much the next dollar is likely to bring (marginal return).
Its limits are real. It looks backward, it is coarse, it struggles when channels always move together, and it says little about a single short campaign. Experiment results are what keep it honest, by calibrating the model where a test has measured the truth.
When the methods disagree
They will. A channel can look strong in attribution and add little in a test, often because it reaches people who were going to buy anyway. When that happens we show both results side by side and decide where to test next, instead of picking the number we like. Putting the methods together for one decision is triangulation.
Rules we measure by
Define the effect
Say which effect, on which outcome and over which period, before launch.
Rules before launch
Windows and success metrics are agreed in advance, and kept the same for similar goals.
Deduplicated against real orders
Conversions are matched against real orders and revenue, so no sale is counted twice.
Platform numbers are a starting point
Experiments confirm them or correct them.
Size the test first
The test is designed before spend, and the control stays clean.
Each method has its job
Experiments calibrate the model, the model guides the budget, attribution guides the day.
Repeat
Results change with competitors, seasons and creative.
Say what isn't measured
Media left out of a measure, and models used to fill gaps, are named.
Show disagreement
When methods disagree, we say so and decide where to test again.
How we run it
We plan the measurement before the media, with the decision, the windows and the tests written down first, and the methods checked against each other. The full approach is on Measurement. The signal it rests on is in our guide Server-side tracking, explained, and the money side in Paying for media across ad platforms.
Not sure your numbers answer the right question?
Send us how you measure today, and we'll tell you which questions it can answer and which it can't.