Most marketing analytics exists to survive meetings. Dashboards nobody opens between reviews, attribution models everybody quietly distrusts, and quarterly readouts engineered so no decision is ever forced. I build measurement that would survive being explained out loud, to your CFO or to the customer being measured.
Tracking gets audited before it is trusted, because the setup nobody checks is where numbers go wrong first. Attribution is treated as a set of useful lies to be triangulated, with phone calls and offline outcomes pulled back into the picture your dashboard currently ignores. Incrementality gets proven with holdout tests any marketing team can run, no data science department required; I validated a 342-to-1 return this way when the dashboard's version of events was too good to sign my name to. Marketing mix modeling runs where spend and channel spread justify it, with a clear-eyed no where they do not. And governance, the deeply boring UTM and naming discipline, decides whether next year's questions are answerable at all.
The deliverable standard comes from the best client I ever had. Every analysis ends with a decision, with the number that would change it stated in advance. Dashboards get rebuilt until the people they are for actually open them, which usually means fewer charts, plainer language, and one page that says what to do Monday.
Privacy is a design constraint here: consent-gated measurement, first-party data used as a verb, clean rooms fed only with first-party data clean enough to match, and regulated-category setups that have already survived HIPAA scrutiny. If your numbers feel confident and your decisions do not, the gap between them is the engagement. It is usually shorter than the last three vendor pitches you sat through, and it ends with numbers you will defend without flinching.
Get numbers you will defend out loud
The gap between confident dashboards and confident decisions is the engagement.
Start the conversationQuick answers
Do we need a data science team to run holdout tests?
No. A holdout compares a group that sees the ads with a matched group that does not, and a marketing team can design and read one with the right sample size. The 342 to 1 eBay return was validated this way.
Which attribution model should we trust?
Use several and trust none of them alone. Platform reports, analytics, and multi-touch models each see part of the picture, so the work triangulates them, adds calls and offline sales, and settles the questions that matter with holdout tests.
Does measurement still work under consent rules and HIPAA?
Yes. Tags fire with consent, first-party data does the work it is permitted to do, and setups for regulated categories have already held up under HIPAA scrutiny.
From The Signal
View all notesImpressions Are Not Created Equal. Frequency Should Not Be Either.
A search impression, a feed impression, and a banner impression are three different events wearing one name. Why every channel needs its own frequency math.
June 27, 2026Most Attribution Dashboards Get Built and Then Ignored
A dashboard nobody opens has a trust problem. The fix is admitting uncertainty on the dashboard itself instead of hiding it behind false precision.
April 3, 2026Marketing Mix Modeling Came Back From the Dead Because Nothing Else Survived
MMM was the dusty econometrics of big TV advertisers until signal loss made it modern again. What mix modeling does, what it needs, and who should wait.
April 1, 2026You Do Not Need a Data Science Team to Run a Holdout
Incrementality testing sounds like a research department. It is a spreadsheet, a map, and the discipline to leave some money on the table for six weeks.
March 25, 2026Every Attribution Model Is Wrong. Choose the One Wrong in Your Favor.
First-touch, last-touch, linear, data-driven: what each marketing attribution model assumes, what it hides, and how to use models without believing them.
March 24, 2026Your Match Rate Is the Ceiling on Everything Else
Every first-party data strategy depends on one unglamorous number: how much of your customer file the platforms can recognize. Typical is about 20 percent.