Marketing analytics
The tracking plan comes first
Dashboards built on bad instrumentation are confident and wrong, which is worse than having none.
Define, in writing, before implementing: every event, when it fires, its properties and their types,
and the question each one exists to answer. An event with no question behind it is noise that will
be maintained forever.
Naming convention decided once and enforced: object_action, lowercase, past tense. Inconsistent
naming is unfixable later without breaking historical data.
Auditing existing tracking
Numbers nobody trusts usually come from one of:
- Double-firing on route changes in single-page apps.
- Events that stopped when someone changed a selector or a component.
- Definition drift — two tools counting "signup" at different moments.
- Bot and internal traffic never filtered out.
- Consent and blockers removing a meaningful and non-random share of data.
Verify by doing the action yourself and watching the event arrive with the properties you expect.
Not by reading the dashboard.
Attribution
Every model is wrong in a known direction. Pick deliberately and state the bias:
- Last-touch — over-credits closing channels: brand search, retargeting. Under-credits
everything that created demand.
- First-touch — the mirror image; over-credits discovery.
- Multi-touch — better, and dependent on complete tracking you probably do not have.
- Incrementality testing — the only method that answers "would this have happened anyway." The
most expensive and the most trustworthy.
Use one model consistently for decisions, and check it periodically against a holdout. Switching
models to make a channel look better is how organizations mislead themselves.
Reporting
Every report answers one question for one audience. Reports built to display everything get read by
nobody.
Show the metric, its comparison period, and the decision it informs. A number with no comparison is
not information. Where a number moved, the report should say why or say that the cause is unknown —
"unknown" is a legitimate and useful finding.
Sources
references/sources.md in this skill lists the outside authorities that settle the questions
here — what each one is authoritative for, and what you may do with it. Check them before
answering on anything they cover, and cite what you used. Most are free to read and not free
to reproduce; the use note on each is binding.
Product and web analytics: Google Analytics 4, Amplitude, Mixpanel, PostHog, Plausible,
and similar.
Attribution: HubSpot or Salesforce campaign reporting, Dreamdata, Rockerbox, and similar.
All of them model rather than observe — treat the output as directional and say so when
you present it.
Warehouse-native reporting beats a vendor dashboard the moment you need to join spend to
revenue on your own definitions.
Never
- Add tracking before the plan names the events and their properties. Retrofitting a schema onto live data is a migration, not an edit.
- Report an attribution number without saying which model produced it. The same period looks like different businesses under first and last touch.
- Change an event definition without versioning it. Every historical comparison silently becomes wrong.
- Build a dashboard nobody named a decision for.
anthropics/skills180k
