Audit an ad account after a bad week

Four things can manufacture a bad week out of nothing, and all four are cheap to rule out.

A partial day. A range ending today is reporting a day that is not over. Comparing it against whole days is the easiest bad week there is to invent.

A restatement. The window you are reading may not be the window you read last week. Platforms hand back revised figures for days that are already stored, a sync takes them, and a sync that rewrote a lot of that history marks the connection restated. Check for that mark before concluding a period moved on its own — a restated fortnight explains itself.

A sync that did not finish. A connection whose last sync finished with errors under-reports, because some accounts landed and some did not. A connection that has never synced answers every reading question with nothing at all, which looks like collapse and is not.

A blended estimate read as exact. A total spanning several currencies is converted at published rates into whichever currency carries the most spend and labelled an estimate, carrying the date of the oldest rate it leaned on. Where a rate is missing nothing is invented: the blended total is refused, and a list that cannot be ranked comparably comes back in raw order with that fact attached, because a raw order across currencies is not a ranking.

One more reading rule that turns a flat week into a catastrophic-looking one: a ratio with a zero denominator is empty rather than zero. An account with no conversions has no cost per conversion. Reading the blank as a nought asserts that acquisitions were free, and reading it as infinity asserts a disaster. It is neither; it is undefined.

Now the record, in this order.

What changed in the window. Not only what was applied. The log carries the whole life of each change — written, edited, approved, applied, and however it ended — and each row carries what it moved and from what. That is what lets you reconstruct the account's state on any day of the window rather than only its state now.

Which authority cleared each one. A person, a credential, or a rule your workspace wrote. If a rule cleared them, read the rule rather than the rows — a run of rule-cleared changes is a question about the rule's cap and scope.

Whether the approver drafted it. The row says whether the approver was the same actor that drafted the change. A run of self-approved rows from one credential is a finding whether or not any single row was wrong.

Whether it was us at all. Where a platform exposes its own change history, it says who changed what, when, from where, and exactly which fields moved. A change present there and absent from Tempio's log was made outside Tempio. On an account several people touch, this is the single highest-yield check in the whole procedure.

The arrangement that makes this readable at all, stated once: your connections read your accounts, a change is written as a draft on the record with the state it would change from and the checks it passed, applying it is gated on approval — by you, or by a rule you wrote — and no platform write reaches a customer account until that platform has granted write access. Applying a cleared change to a platform is in private beta.

Bad weeks are almost never evenly distributed, and the aggregate hides that by construction.

Which campaigns carried the spend last week, and how does that compare with the fortnight before?

Which campaigns spent the most and produced the fewest results in that window?

Give me the brief for that ad account.

The brief covers the last ninety days with a comparison against the previous twenty-eight, the objective mix, how concentrated the spend is, the rankings by return and by cost per acquisition, and the weekday rhythm. It is computed rather than live and carries its own age, so say when it was computed before quoting it.

The concentration question is the one that decides the response. If most of the damage sits in one campaign, the account's aggregate is that campaign wearing a disguise and the fix is narrow. If it is spread evenly, the cause is usually upstream — a landing page, an offer, a tracking change, a seasonal floor — and no campaign-level change will touch it.

Two reasons to be conservative here, and both are structural rather than temperamental.

Undoing a disruptive change is itself a disruptive change. An undo is a new draft built from the state the entity was in before, going to the same gate with the same guardrails, and the delivery system is asked to adjust again rather than returning to where it was.

And a week diagnosed on partial data plus three simultaneous corrections is a fortnight nobody can attribute. One change, a written prediction, and enough time to read it.

If the diagnosis is that something should not have been applied, the response is a drafted inverse rather than a setting. If the diagnosis is that a rule cleared something it should not have, narrow the rule before you fix the symptom — otherwise it clears the same thing again next week.

And if you are not yet sure, the control that buys time without destroying anything is pausing the connection: it stops every apply written against it, whatever is queued and whatever rule would otherwise have cleared it, and nothing is rejected.

The finding that matters is not what happened but what you will check first next time. Four lines is enough: what looked wrong, what it actually was, what you changed, and what you expected that change to do.

The account audit template is the long form of this procedure as a document, and the performance review template is the weekly rhythm that turns most bad weeks into a thing you noticed on Tuesday instead of a thing you discovered on the following Monday.

Who approved what, and when is pass two in detail, Review a batch of campaign changes is how the changes you are now reading got approved, and Give a contractor limited account access is what to fix if pass two found somebody who should not have been able to do that.