AI for Google Ads: how to brief a task that can actually be finished

"Run an audit of my account" is not a task

"Let's audit the account." "Let's optimize performance." "Check what's wrong." Everyone nods, and then everyone does something different.

These sentences share one property: there is no moment when they are finished. Nobody gets to say it's done, because nothing decides that it is. They are reasons to sit down and work. Nothing in them says when to stand up.

The test is simple. Ask yourself: how will I know this is complete? If the answer is "well, things will be better," the brief doesn't exist yet.

The four fields

A task someone can act on without guessing has four fields. None of them is a formality, and each fits in one sentence.

Goal — what is different when we finish

Phrased so that it can be verified.

• Weak: "better results"
• Good: "every ad in the Brand campaign has all 15 headlines and 4 descriptions"

A business goal is not a task goal. "ROAS above 400%" and "more sales" depend on the market, the season and the competition. The work can be flawless and the goal still missed, or the other way around. Write the business goal down as the reason, then write a task goal you can check off when it's done.

Scope — which campaigns, what period

Without scope you get an average of everything, and the average hides exactly what you were looking for.

• Weak: "on the account"
• Good: "all campaigns with spend, last 30 days"

With products, 30 days is usually too short. Whatever sells less often comes out looking weak when it isn't: the long tail needs 90 days before it shows anything at all.

Boundaries — what stays untouched

The field skipped most often and the most expensive to skip.

• "Budget unchanged"
• "Leave the brand campaign alone"
• "Nothing changes without my approval"

Leaving them empty doesn't mean anything goes. It means someone else guesses where they are.

Deliverable — what you want at the end

A list to approve, or a change already made? Those are two different jobs.

• Weak: "make it good"
• Good: "findings on screen, changes only after I approve them"

Three traps that only surface once the work starts

Words that sound precise and mean nothing

"Products to exclude" means nothing until you name the cutoff. One criterion flags twenty products, another flags two hundred, and both readings are correct.

The same family includes "weak campaigns," "expensive keywords" and "unprofitable ad groups." Each needs a number before anything can be counted.

Two questions in one instruction

"Which products earn" and "is our ROAS inflated by the brand" are two separate analyses. They work at different levels and need different periods: products need the long tail, while the brand split needs whole calendar months so that one good week doesn't tip the result.

Averaged together they produce a report where everything looks fine and nothing follows from it.

Order matters too. Profitability thresholds calculated on numbers inflated by the brand will show products as profitable when it is the brand doing the earning. Separate first, then count.

A threshold named too early

A cutoff picked before you see the distribution hits either three products or three hundred. Sometimes the honest move is to leave it open and set it after the first read, when you can see where the line actually falls.

This is the one field you may deliberately leave empty, as long as it is marked as empty rather than guessed.

The same brief, before and after

Before:

Run an audit of my Google Ads account.

After:

Show me which products earn and which ones should be excluded. Separately, check whether our ROAS only looks that good because we are bidding on our own brand. E-commerce account, spend in PMax and Shopping, 90 days of data. Nothing changes without my approval. I want a list of products to exclude and, separately, the numbers with brand stripped out. I am not naming the profitability cutoff yet; we will set it once I see the spread.

The second brief can be carried out. When it is, you can say whether it was. That's the whole difference.

The method works without AI too

These four fields belong to no one and need no particular tool. They work anywhere someone follows an instruction for you, including when that someone is a person. A junior specialist handed "clean up these campaigns" has exactly the same problem as a model: they have to guess what you meant.

The difference is that a model guesses instantly, across the whole account, and without hesitating. Which is why, with AI, describing the work stops being a formality and becomes the work.