How to test paid social creative and get a decision

On paid social the algorithm does most of the targeting and the creative does the rest. That makes a reliable way of finding out which creative works the highest-leverage thing you can build, and it is usually the thing that does not exist.

Fix the structure before you test anything

Most accounts we take on are running too many ad sets on too little budget. Every one of them sits in the learning phase, none gathers enough signal to optimise, and every test run inside that structure is unreadable.

Consolidation is normally the first performance change, and it costs nothing. Fewer ad sets, more budget in each, so the optimiser has enough conversions in one place to learn from. Only once that is true does a test result mean anything.

Test concepts, not variations

The difference between two headlines is rarely large enough to detect at realistic budgets. The difference between two concepts usually is. A concept is a distinct argument for the product: a different problem, a different audience, a different format.

• Change the angle: lead with the problem, then with the outcome, then with the objection everyone raises.
• Change the format: static, motion, and native-feeling footage behave differently enough to be worth separating.
• Change the hook: on TikTok the first two seconds carry the result, and testing hooks against a fixed body is one of the few small variations that genuinely pays.
• Leave colour and button copy alone until you have exhausted the concept space, which most accounts never do.

Give each variant enough spend to be readable

Decide before launch how much each variant needs to spend before you will look at it, based on your cost per acquisition rather than on a round number. A variant that has spent less than a few times your target cost per acquisition has told you nothing, and reading it anyway is how good creative gets killed early.

Run whole weeks. Audience behaviour on a Tuesday is not the same as on a Sunday, and a test that ends mid-week compares two different mixes.

Agree the stopping rule in advance

Write down, before the test starts, what spend threshold ends it and what result counts as a win. Then hold to it. Stopping early because a variant looks ahead is the single most common way testing programmes generate false wins, and a few of those are enough to make a team stop believing in testing at all.

Every test ends in one of three written outcomes: keep, kill, or iterate. An unrecorded test gets rerun a year later by somebody who was not there.

Plan for fatigue, because it is not optional

Paid social decays faster than search because the audience is not asking for anything. A set of ads that stops working after two months is usually not being beaten by a competitor, it is being ignored by people who have already seen it.

That makes creative supply a rate rather than a number. A useful planning assumption is a fresh batch of concepts every four to six weeks, and more as spend rises, because a bigger budget saturates the same audience faster.

Make sure you can see the results at all

Browser-side tracking loses a meaningful share of events to blockers and privacy settings, and those losses are not random. If a large part of your conversions is invisible, the optimiser learns from a skewed sample and your test results inherit the skew.

Server-side events recover much of that signal, which is why measurement work comes before creative work in our paid social advertising engagements, and why it overlaps with marketing analytics services.