Channel testing and attribution

GA4 if you want free and deep; Plausible or Fathom if you want simple and EU-safe.

The six moments your funnel turns on, named once and never renamed. Subscription: signup_started, signup_completed, activated, trial_started, subscribed, cancelled. One-off purchase (a kit, a template, a course): product_viewed, checkout_started, purchased, downloaded_or_installed, first_use, refunded. Usage-billed or free-with-upsell: pick the six that mark the same turns. The names matter less than never changing them.

Go through the funnel as a stranger in incognito — sign up, or buy the cheapest thing — and watch all six land. Most tracking is broken.

One page: source, medium, campaign casing. Inconsistent UTMs make Phase 3 unreadable.

The only free source of truth for what you rank for and what Google refuses to index.

One console for the second-largest reach you have: Bing’s index also serves Yahoo, DuckDuckGo, Copilot and most of Ecosia. It no longer feeds ChatGPT, which runs its own crawler — but it is the only console that reports your Copilot citations.

The single in-product moment after which people stick. Everything in Phase 1 aims at it.

Even if they are zeros. You cannot claim improvement without a starting line.

A free-text box beats any tracking script for dark-social discovery. Read every answer.

Use the convention you wrote in Phase 0. Untagged links become “direct” and hide your winner.

First touch tells you what to invest in; last touch tells you what to fix.

The channel with the most signups is often the one with the worst retention. Score channels on people still here at day 30.

Five rows: channel, signups, cost, CAC, trend. Anything longer gets skipped by week three.

Sessions rank your posts one way and signups rank them another. Carry it through to activation and retention too — pieces that pull the wrong reader show up as churn months later, not as a bad traffic number.

GA4 files ChatGPT, Gemini, Copilot and Grok referrals under AI Assistant already — but not AI Overviews or AI Mode, which stay inside Organic Search. Judge the row on conversion rate, not sessions: it is the one place on your own property where “the assistant pre-qualified them” is checkable, and it counts the clicks only, never the reads.

App Store links take a campaign token (pt and ct) and Google Play reads UTM source and campaign on store links; both report in the store’s own analytics. Apple lists a campaign only after 24 hours and shows a figure only once it reaches five.

Where your ICP already gathers, not where you are most comfortable posting — and count the channel Phase 3 built as one of the five, scored on the same card as the newcomers.

Volume, budget, and the number that counts as a pass. Written before you start, not after.

Parallel, not sequential — sequential testing takes five months and the market moves.

Ceiling matters most: a channel that works but caps at 20 signups a month is not the winner.

Price your own hours in. A channel that is free only because your time is unpriced is not free.

The phase exit is one channel at 40% or more. Below it you have not found the winner, you have a spread.

The sunk cost is already sunk. Half-running four channels is how companies stall here.

Appear on somebody else’s podcast, channel or newsletter in your space and watch what their audience does. If it is one of their better performing pieces, that is a real signal for a fraction of the cost of finding out yourself.

Before you optimise anything. Volume finds the ceiling faster than cleverness does.

If it lives only in your head, it cannot be delegated in Phase 4 and it dies when you get busy.

Consistency beats intensity in every distribution channel that compounds.

Support load, moderation limits, ad auction, your own hours. Know it before you hit it.