About
About this site
Learn what marketing data can tell you, what it cannot, and how to make better decisions when tracking, attribution, and experiments are imperfect.
What this is
A field guide to marketing measurement when the data is imperfect, which is always. It covers what marketing data can tell you, what it cannot, and how to make a defensible decision in the space between.
It is organized around decisions rather than algorithms. A method appears here only when it is attached to a question someone actually has to answer, and it arrives with its assumptions, its diagnostics, and the conditions under which it stops working.
What it is not
Not a catalog of models, and not a portfolio. There is no attempt at exhaustive coverage of marketing methods. The bet is that a small number of subjects treated honestly and completely is worth more than a broad index of shallow definitions.
Where a topic is on the roadmap but unwritten, it is labeled as planned and does not link anywhere. That is intentional.
Who writes it
Reid Rhodes, a marketing data scientist working on measurement, incrementality, and growth decisions, based in Chicago, IL.
The work behind this site has mostly been measurement for programs spending at eight figures a year: experimentation, attribution, and incrementality across B2C ecommerce and B2B SaaS, plus the data engineering underneath when the pipeline was the reason nobody trusted the number. Before that, six years advising clients on money, which is a useful apprenticeship in explaining uncertainty to someone who wants a straight answer.
Most of what is here comes from having to answer these questions under real constraints. Tracking that lost a third of conversions to consent. A campaign that launched three days late into the wrong audience. A CFO who needed a number on Thursday. The methods matter, but the judgment about which method survives contact with your actual data matters more, and that part is harder to find written down.
How the examples are sourced
Every worked example declares its data provenance at the top of the page: public, simulated, or sanitized. Simulated examples are generated by seeded scripts committed alongside the writing, so the numbers in the prose, the figures, and the frontmatter are regenerated together and cannot drift apart. A build check enforces that.
The source repository is private. If you want to read the estimators or walk through how a dataset was generated, ask and I will show you.
Simulated results are never presented as a real company's outcomes. Where an example is based on real work, the business, the absolute volumes, and identifying details are changed, and the page says so.
A note on certainty
The site tries to model the behavior it argues for. Estimates appear with intervals. Assumptions appear with the consequence of their failure. Every worked example ends with a section on what it does not establish. If that reads as hedging, it is not. It is the difference between a measurement program that survives scrutiny and one people quietly stop trusting.