Measurement
Attribution guides the day, experiments show what the media caused, and the mix model sets the budget.
At the end of a quarter we lay the platforms' reports next to the orders the business actually booked. Each platform's number is honest about what it saw. Added together, they often claim more sales than there were.
That gap is where our measurement starts. Attribution tells you where conversions happened. It can't tell you what caused them, so we build measurement in layers and check each layer against the others.
Each method answers a different question. Used together and checked against each other, a practice called triangulation, they give one answer you can act on.
Where did conversions happen? It guides daily optimization inside and across digital channels.
What did the media cause? Holdouts, geo tests and matched markets compare people or places that saw the campaign with those that didn't.
Where should the next budget go? A model of sales against spend in every channel, digital and traditional, with no tracking of individuals.
Calibration Experiment results calibrate the mix model. The model sets the budget, and attribution steers the day inside it.
Platform-reported results are where we start, not where we stop. When attribution says one thing and an experiment says another, we show both, explain the gap and test again where the doubt sits.
Search, social and the open web are measured with clicks, conversions and lift tests, as set out on Digital advertising. Many channels never get a click. Each one leaves its own trace, and we measure it with that trace, then bring it into the mix model and geo tests with everything else.
Rating points, reach and frequency for linear TV, deduplicated with streaming where the data allows. Household impressions and completion from the servers that insert ads into streams, site visits right after an ad airs, and outcome studies.
Radio audiences from the measurement the medium is traded on, then geo tests, brand lift, promo codes and dedicated URLs, and impressions counted on each file request where ads are inserted dynamically.
Audiences from the industry's audience measurement, as opportunity to see and likelihood to see, proof of play on digital screens, and exposed areas compared with control areas to read the lift.
Audited circulation, promo codes and dedicated addresses, matched markets where print runs region by region, buyers matched back against the mailing list, and control groups that receive nothing.
Closed-loop sales from the retailer's data, with a stated model for the sales that can't be matched, such as cash or unrecognized cards.
Codes, registrations, surveys, and test and control areas. The effect is hard to separate from the rest of the plan, so it is reported as an estimate.
Modeled store visits, out-of-home audiences and any channel left out of the model. Each one is labeled as an estimate in the plan and in the report, with the method behind it.
The one your data can support, usually data-driven where there is enough volume, with windows as close as each platform allows for similar goals. We read it as a guide for daily decisions. Proof of cause comes from experiments.
Some effects are too small for the test a budget can afford, and some channels only report estimates. The plan and every report name them, with how the gap was handled.
Big enough to detect the smallest effect worth acting on, the minimum detectable effect. We size it before spend, from that effect, the noise in your sales and the time the test can run. If a budget can't support a clean test, we say so before it starts.
Bring a plan, a report or a question your numbers can't answer yet.
