How to Structure a Creative Testing Campaign From Scratch

Before Ads Manager, write one sentence: what you are testing, what stays identical, and what decision you will make at 72 hours. Example: 'Which of five hooks produces the lowest cost per purchase with the same avatar, offer and landing page.' If you cannot name the one variable, you do not have a test yet — you have a pile of ads. Tape that sentence above the campaign so mid-week 'improvements' have to beat it in writing.

Isolation is what makes the spend diagnostic. If variant three also changes the avatar and variant four also changes the discount, a winner tells you nothing you can reuse next week. Hold product, offer, destination, length, captions and persona constant. The most common contamination is a 'small' mid-test edit to the supposedly locked body script because one line felt off — that edit makes the whole round unreadable. Write the locked body somewhere visible and do not touch it until the test is called.

Decide the success metric up front, in writing: cost per purchase (or your true bottom-funnel event), with hook rate and click-through as supporting evidence, not the other way around. Optimising the test for clicks finds people who click, and click-lovers systematically do not buy. Rank by cost per result at the end; use hook rate (3-second views divided by impressions) to explain why a video died or lived. If you cannot name the bottom-funnel event, you are not ready to spend on a test yet at all.

Generate a structured batch: one product, three genuinely different hooks times two avatars, which gives you six videos. Add two statics only if you already know you will need them for retargeting later — they do not belong in the cold creative tournament. Random one-off videos cannot tell you whether the hook or the face did the work. Hold offer, length and landing page constant across the six or the matrix stops isolating anything.

The matrix is diagnostic because every hook appears with both avatars and every avatar carries all three hooks. If hook two wins with both faces, the hook is the driver and the next round iterates messaging. If avatar A wins across all three hooks, the face is the driver and the next batch explores similar personas. Six unstructured videos cannot separate those effects. Keep the three hooks on different angles — problem, social proof, curiosity — not three rewordings of the same opening line. If two hooks are the same sentence with a different adjective, you have already wasted a cell.

Klip Kanvas is the fast way to fill that matrix from one product brief: lock the body script, generate the three hooks, render each hook on two avatars, and export 9:16 plus 1:1 so placements are not a hidden variable. Hold video length constant across the six. If you cannot afford six, drop to three hooks on one locked avatar rather than shrinking spend per ad below the point where you will get a read. A short, fully isolated three-cell test beats a six-cell test you underfund halfway through the week.

In Ads Manager, create a new Sales (conversions) campaign. Do not use Advantage+ shopping or Advantage+ creative for the test. Use a single ad set, broad targeting, automatic placements with Audience Network excluded if you want a cleaner creative read, and drop all six videos in as separate ads — not as one flexible creative unit. Confirm the conversion event is Purchase (or your real close) before you hit publish.

Consolidation is arithmetic. Meta's learning phase wants roughly 50 conversion events per ad set per week before delivery stabilises. Split a testing budget across six ad sets and each one crawls toward that threshold for a month; pool it in one ad set and you can cross it in days. The auction already runs a creative tournament inside the ad set. Your job is to feed that tournament enough data to finish, not to referee it from six starved ad sets. If your weekly purchase volume cannot support even one ad set's 50 events, lower creative count, not ad-set count.

Leave targeting broad on purpose. For UGC video, the creative is the targeting; interest stacking adds a second variable you cannot isolate. Geography and age stay as operational constraints (where you can fulfil, where unit economics work), not as a guess about the customer. Bid for purchase, not for landing-page views. ABO versus CBO barely matters with one ad set — pick ABO so the test budget cannot be reallocated away if you later add a second ad set by mistake. Every extra ad set you add 'for control' is another place learning can stall.

Set a daily ad-set budget of roughly two to three times your target CPA, which is enough to rank six ads in 48–72 hours without pretending you can fine-read them. Equal delivery across ads matters more than a large total. Do not pause 'losers' at hour six — the learning phase needs the full window. Put a calendar hold on Ads Manager for those 72 hours if you know you will otherwise tinker.

Work backwards from the decision. To kill a creative you want about 1,000–2,000 impressions on it; to scale one you want several conversions attributed to it. Across 72 hours you should be able to rank the six, not to write a dissertation on each. That is the test's job. Plan the week so total testing spend is in the neighbourhood of 20–30 times your target CPA, which is enough for a clean batch without turning the lab into your only prospecting campaign. If the daily budget cannot support 1,000 impressions per ad, drop a variant rather than starving all six.

The no-touch rule exists because every meaningful edit resets learning. Pausing an ad, changing budget by more than about 20%, or editing copy sends the ad set back into learning, and the two days of data you had stop predicting the next two. Schedule the campaign to start at midnight in the ad-account timezone so day one is a full day, and spend the 72 hours preparing the next matrix instead of watching this one. Checking the dashboard is harmless; editing bids, copy or budgets is not, and it is the edit that costs you the window.

Use a file-and-ad name that encodes product, hook, avatar and ratio — for example serum_hook-proof_avatarA_9x16. Put the same string in the ad name in Ads Manager. If you cannot reconstruct the variable from the name, the test will teach you nothing when you export the CSV. Do this before you upload, not after the first report looks messy and you cannot remember which file was which.

Naming is part of the structure, not admin. Teams that skip it end the week knowing 'ad 3 won' and not knowing whether ad 3 was the question hook or the complaint hook. That is how accounts rerun the same test for months. Keep a simple sheet with the hypothesis, the six names, go-live time, daily budget, and the decision rule. When you call the test, write the winner and the why in the same row so the next matrix starts from a documented angle, not from memory. Export the Ads Manager breakdown at 72 hours into that same sheet so the names, spend and CPA live in one row.

At 24 hours, audit delivery only — is any ad stuck at near-zero impressions for policy reasons, and is spend roughly even. At 48 hours, compare hook rate and CTR. At 72 hours, rank by cost per result, with hook rate as supporting evidence. Do not crown a winner on day one. Early CPA on a handful of purchases is not a ranking; treat it as a delivery check until the window closes.

A hook rate of 30% or better on cold traffic is a healthy first-three-seconds bar; CTR around 1% or better on cold is a healthy click bar. Read them in funnel order: hook rate diagnoses the open, hold rate (15-second views divided by 3-second views) diagnoses the middle, CTR diagnoses the close, and CPA/ROAS diagnoses whether the click was qualified. A strong hook with a weak CTR needs a new CTA, not a new first line. A poor hook is dead — replace the first two seconds and relaunch that slot. A video that wins on CPA but loses badly on hook rate is often a tracking or audience accident, not a hook you should scale.

Do not crown a winner on fewer than three or four conversions, and treat a single outlier purchase on tiny spend as noise. A variant that wins on cost per result by at least 15 to 20% over second place, with enough spend to trust it, is your winner. If two ads are within about 10% of each other, that is a tie: extend 24–48 hours or keep both as co-winners. A hook-rate gap of five percentage points or more at 48 hours is usually real, but conversion still needs the full 72 hours. Write the ranking in the sheet at 72 hours even if you keep both co-winners — the record is the point.

After 72 hours, kill anything that spent more than 1.5× target CPA with no conversions. Iterate middles that had a strong hook and a weak close. Duplicate clear winners into a separate scaling campaign. Leave the test campaign standing as the permanent laboratory for the next matrix. Resist the urge to 'just bump' the winner where it sits; that bump is a scaling decision and it belongs elsewhere.

Run three buckets. Clear kills: poor hook rate or spend past the kill line with nothing to show — pause them and write down which hook style or avatar lost. Fixable middles: strong top-of-funnel, weak close — keep the proven two seconds and regenerate the body and CTA. Winners: at or better than target CPA with multiple conversions — they leave the lab. Scaling inside the test campaign contaminates the next round and invites budget changes that reset learning. Harvest the lesson from kills on the same day you pause them or you will rerun those hooks next month.

The next test changes one new variable. If the hook was the driver, lock that hook and test avatars or offers. If the face was the driver, lock the persona and test new hooks. Do not reward a win by dumping new audiences into the same ad set. Horizontal learning happens in the next matrix; vertical spend happens in the scaling campaign, where you will raise budget 20–30% every 48 hours rather than doubling overnight. The lab's only remaining job after a call is to host the next matrix, not to carry last week's winner at 4x spend.