A week of ad changes on the record

The first question of most weeks is whether anything moved while nobody was looking.

With a record, that is a lookup. Every applied change is a row carrying what it moved, from what value, when, and which authority cleared it — you, a credential you minted, or a rule you wrote, named. A quiet weekend produces no rows, and no rows is an answer rather than an absence of information.

Without one, the question has no cheap form. You compare what you remember against what you see and conclude that nothing much changed, which is correct most weeks and expensive the week it is not.

Somebody asks for a change and what comes back is a group of proposals rather than a single one.

Reading the group before the rows is what saves the afternoon. The count against what was asked for is the first signal: a request about two campaigns that produced fifteen proposals is not fifteen things to read carefully, it is one question about what the instruction was understood to mean.

The rows themselves carry what they would change from, so a figure can be judged without looking anything up. And any row near a limit says so — an amount close to the account's own ceiling, a large move from the current value, a budget shared across several campaigns so that changing it changes all of them. A warning changes no value and refuses nothing; it tells the reviewer what is near, which is the information a reviewer actually wants.

A rule you wrote cleared something overnight. The record says which rule, by name, on which row, against which previous value.

This is the case people worry about most and it is the one the arrangement handles best, provided two things are true. The rule can only clear the reversible half of the work — pauses and decreases, under a ceiling it states, no further than a distance it states — and the record distinguishes a change a rule cleared from a change a person approved.

The second is easy to underestimate. A system that flattened both into approved would be telling you the change was wanted without telling you by whom, which is precisely the distinction the record exists to preserve. There is a related detail on the same row: whether the thing that approved a change was the same thing that proposed it, because a credential approving its own work is a different fact from a person approving it.

Every row in that week rests on the same arrangement. Reading is what a connection does; changing is not. A change becomes a row before it becomes a fact, carrying the value it would replace and the checks it survived, and it moves only when an authority clears it — you, or a rule you wrote. Past that gate, no platform write reaches a customer account until that platform has granted write access, and the step that reaches a platform is in private beta.

Something did not happen, and the useful version of that is a reason rather than a silence.

A proposal can be refused for ordinary reasons that are worth knowing before you draft again: a daily figure above the ceiling the account's own recent spending supports, a move larger than one change is allowed to make, a run of increases on one connection already adding up over a rolling day, or a currency that does not match the account it was aimed at. Each refusal names the number.

That last class is worth its own note. The check that sums applied increases over a rolling day is the only one that can refuse something every earlier check allowed, and it exists for exactly the case this post is about: ten individually reasonable raises across a week are still a doubled account, and no single row looks wrong.

The end-of-week question is not what the numbers did. It is whether what the numbers did has anything to do with what anybody changed.

With the record, that is a comparison rather than a theory: the rows from Monday to Thursday, against the daily counts either side of each date. The answer is frequently that the changes had nothing to do with it, and establishing that in ten minutes is what makes the rest of the afternoon useful.

Two habits make the comparison honest. Compare like windows — a range ending today includes a day that is not over, and the resulting dip is an artefact of when you looked. And check whether the platform restated the period, because a stretch of stored history that was rewritten explains a move that nobody made.

Where a platform exposes its own change history, that is the other half: it says who moved which field and when, so a change present there and absent from your own record was made outside the system entirely. On an account several people touch, this is the single highest-yield check available and almost nobody runs it.

The same five questions, with the same answers eventually, arrived at by memory and inference instead of lookup.

The cost is not mainly the time. It is that inference produces confident wrong answers, and confident wrong answers about what changed become the input to the next decision. A quarter built on four of those is a quarter nobody can explain, and the explanation is usually demanded at the least convenient moment.

A record is not for the auditor. It is so that the person who runs the account can stop remembering and start looking things up.

Reading a change request like a diff is the skill this week assumes, and Audit an ad account after a bad week is the procedure for the week that really did go wrong.