How to set up alerts across client accounts, and keep them worth reading.

It is set to fire on movement, and marketing numbers move. Spend, cost per result and conversions all vary day to day for ordinary reasons, so an alert on change fires constantly, and within a fortnight somebody routes it to a folder they do not open.

Across a roster the arithmetic gets worse. A rule that fires once a fortnight for one client fires most days once it runs for thirty, so the bar has to be set with the whole roster in mind.

The bar that keeps a channel alive is narrow. An alert should mean that a named person has to do something today. Everything else is a chart, and a chart is something you look at when you choose to.

In this order, because the first one is the most valuable and the least often configured.

• Data that did not arrive

  A sync that failed makes every number downstream wrong while the dashboard still renders normally. Nothing looks broken, so nobody checks, and the report goes out with a partial month in it.

• A value outside a band you set

  Spend above a client's cap, cost per result above what the account can carry. You are alerting on a limit somebody already agreed with the client, which is why this one rarely gets muted.

• A change against the same period

  This week against the same week, not against yesterday. Day on day comparison alerts on the weekend, on paydays and on every campaign that runs to a schedule.

• A number that did not move at all

  A metric sitting at exactly the same value for several days is usually a stuck feed rather than stable performance. It is the earliest signal you can get, and almost nobody configures it.

Start from the action, not from the number. Ask what you would do in the next hour if this fired. If the honest answer is that you would look at it later, it is not an alert and putting it in the channel costs you the alerts that are.

Then set the level from the metric's own history rather than from a round number. Look at how the figure has moved over the last few months and put the line outside the range it lives in normally. A threshold picked because it is a tidy figure fires on ordinary variation, and every time it does it teaches somebody to ignore the next one.

One owner per alert, named. A channel that everybody is in is a channel nobody is responsible for, and the failure mode is not that it is missed but that each person assumes another has picked it up. For an agency, the owner is usually the person who runs that client's account.

The message needs enough to act on without opening anything: which client, which metric, the value, what it is being compared with, and a link to the place the decision gets made. An alert that only says a threshold was crossed starts an investigation rather than ending one.

Checks run against the modelled tables, per client, as four rule types: a threshold, a percentage change, a comparison against the same period, and data simply not arriving.

Because they read the same tables the dashboards read, an alert and a report cannot tell you different things about the same day. The freshness case is the one worth setting up first, since it is the failure that a dashboard is least able to show you.