Set attribution thresholds worth alerting on

A threshold is only worth having if crossing it triggers an action you decided on beforehand, and if the crossing is bigger than your own uncertainty. Most fail both tests.

Build it action-first

Starting at step four is the common error, and it produces a number with no attached behaviour. A threshold nobody has decided to act on is a fact about a spreadsheet.

Use the interval, not the midpoint

This is the single change that makes thresholds usable. A midpoint crossing a line means very little when the confidence interval spans both sides of it. A whole interval sitting on one side is a genuine statement.

So express thresholds as interval conditions. "Upper bound below break-even" is a cut signal. "Lower bound above break-even" is a scaling signal. "Interval straddles break-even" is neither, and naming it explicitly as a third state is what stops teams forcing a decision out of an unresolved read.

Three states, not two. The report structure that carries them is in which channels to cut, for the CFO.

Break-even is not one number

Your break-even depends on margin and on which costs you include. Write down which definition you are using, because two people in the same team will otherwise use different ones and the threshold becomes a source of argument rather than a resolution to one. The true ROAS guide covers the arithmetic.

What to do about the middle state

Stage the change. If the interval straddles break-even and the channel is large, reduce by a smaller amount and re-read after a full window rather than making the full cut or none of it. That is a real option and most threshold schemes have no room for it.

Set them before you look

Thresholds decided after seeing the data are conclusions with a threshold drawn around them. Agree them in a meeting, write them down with a date, and treat changing them as a decision that itself needs a reason.

Agreeing them in advance is also what stops the numbers being relitigated later, which is discussed in the lift number your agency will accept.

What we return

Every channel gets a data-health score from 0 to 100 that grades the read, not the channel, and a next step: raise, cut or hold, or fix tagging first. From a Google Analytics export, €99 for a first read, refundable within 30 days, no questions asked.

Related answers

• Own your alert rules, not the vendor defaults
• Real-time attribution alerts during peak trading
• The cost of reacting to attribution noise
• Continuous reads let you test more often