Why Creative Is the Only Lever Left

Ten years ago a media buyer earned their fee through targeting. Interests, lookalikes, layered audiences, exclusions. The craft was in the setup, and skill there produced a real advantage.

Almost all of that has been absorbed by the platforms, and the businesses still spending their effort there are optimising a lever that has been disconnected.

What the platform now decides

In December 2024 Meta published the engineering detail behind Andromeda, the retrieval system that determines which ads are eligible to be shown to a given person.

The reported figures are worth reading for what they imply rather than for the numbers themselves: a ten thousandfold increase in model complexity for personalisation, a six percent recall improvement to the retrieval system producing an eight percent ad quality improvement on selected segments, and a twenty two percent increase in return on ad spend among advertisers who enabled Advantage+ AI features.

The mechanism matters more than the percentages. The system does not begin from an audience you defined. It evaluates the creative, the copy, the format and historical engagement, then predicts who is likely to respond.

Which means your creative is now the targeting. The asset itself is the instruction, and the audience is an output rather than an input.

Three variations of one image is not a test

Most accounts describe themselves as testing creative and are doing nothing of the sort.

Changing the background colour, cropping differently or swapping a headline word produces variations of a single argument. The system learns almost nothing from that, because everything that determines who responds has stayed the same.

A genuine test runs different arguments. One about cost, one about time, one about risk, one about what happens if nothing changes. Those reach different people, because the system finds the audience that responds to each, and the results diverge in ways that actually tell you something.

What a real angle looks like

The useful distinction is between what you are saying and how it looks.

An angle is a claim about the buyer’s situation. You are paying two people to do work that does not need a person. Your best leads are contacting three competitors while you are on site. You are quoting jobs you were never going to win.

Format is whether that arrives as a static image, a short video, a carousel or plain text. Both matter, and they are separate experiments. Businesses that confuse them conclude video does not work when what actually failed was the argument inside it.

Decay is faster than most people plan for

Creative performance falls predictably, and on a smaller audience it falls quickly.

Frequency climbs, response drops, cost per result rises gradually rather than suddenly. There is no day it broke, which is why it is usually noticed a month after it started.

That makes this a supply problem rather than a strategy problem. Something new has to be entering the account continuously, and for most businesses the constraint is production capacity rather than ideas. A design queue that turns around four assets a month cannot feed a consolidated account.

Where AI removes the constraint

Two jobs, and together they are what makes a modern account work.

Volume. AI produces variations on an angle at a rate no queue could sustain, so the account always has something entering it rather than running three assets until they are exhausted.

Judgement about what worked. More usefully, it reads what happened downstream. Which angle produced held calls and closed deals, using data from your CRM rather than the platform’s attributed numbers. That is the feedback loop that decides where budget goes, and it only exists when the ad account and the CRM are genuinely connected.

Without the second, volume just produces more noise. With it, the account gets better at finding the argument that sells rather than the one that gets clicks.

Where the angles come from

Not from a brainstorm. From things people have actually said.

The objection your salesperson hears every week. The sentence a client used when they explained why they finally moved. The reason somebody gave for choosing a competitor. The question that comes up in every first meeting.

Those are angles that already work in conversation, which is considerably better evidence than anything produced in a room by people describing what they think the market thinks.

What to measure

• Number of genuinely distinct angles live, not number of assets.

• Frequency by ad set. Rising frequency with falling results is decay, not a targeting problem.

• New assets entering per month, against what the account consumes.

• Cost per closed deal by angle, which requires the CRM connection and is the only number that says which argument sells.

Build a library rather than a campaign

The businesses that keep an account fed treat creative as a standing asset rather than something produced per campaign.

That means a stock of raw material: photographs of actual work, short clips from site, screenshots of real output, quotes from things clients have said. Gathered continuously rather than commissioned when an account runs dry.

Most businesses have far more of this than they realise, sitting on phones and in project folders. The constraint is not access, it is that nobody has been collecting it deliberately, so when creative is needed the only option is a stock image and a headline.

Real beats polished more often than expected

One consistent pattern worth knowing before spending on production.

Material that looks like it came from inside a business frequently outperforms material that looks like an advertisement. A photograph taken on site, a short unpolished clip, a screenshot with real numbers on it. These read as evidence rather than as marketing, and evidence is what a sceptical buyer is looking for.

That is convenient, because it means the constraint on volume is not budget. It is whether somebody remembers to capture what is already happening.

Keep the losers

An angle that failed is information, and most accounts discard it.

Record what was tested, what it argued, and how it performed against closed deals rather than clicks. Six months later that record prevents you retesting the same argument in a slightly different jacket, which is otherwise exactly what happens when the person running the account changes.

Sources

• Meta Engineering, “Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine”, 2 December 2024.

AI Optimize produces the creative volume a consolidated account needs and connects it back to closed revenue, so you learn which argument sells rather than which one gets clicks. That work sits under Paid Ads Management.