How-to Launch High-ROAS Ad Campaigns with AI Creatives

Before touching Ads Manager, generate a proper test batch in Klip Kanvas: one product, 3 hooks × 2 avatars = 6 videos, plus 2 static ads for retargeting. Testing one creative tells you nothing; testing a structured batch tells you which hook, which face and which angle your audience actually responds to.

The matrix structure is what makes the batch diagnostic rather than decorative. Because every hook appears with both avatars and every avatar carries all three hooks, any performance gap isolates cleanly to one variable. If hook two wins with both avatars, the hook is the driver and you iterate on messaging next round. If avatar A wins across all three hooks, the face is the driver and your next batch explores similar creator profiles. Six random videos can't tell you any of this — six structured ones can.

Keep the batch honest by holding everything else constant: same product, same offer, same landing page, same video length. The most common testing mistake is changing the discount or destination between variants 'while you're at it', which contaminates every read. Your three hooks should also attack genuinely different angles — one problem-led, one social-proof-led, one curiosity-led — rather than three rewordings of the same idea, or the test will return a shrug.

Create a single campaign with Advantage+ (Meta) or Smart Performance (TikTok) turned off for the test phase. Use one ad set, broad targeting, and drop all 6 videos in as separate ads. Modern ad platforms allocate spend to promising creatives automatically — fragmentation into many ad sets only starves the algorithm of data.

The reason consolidation wins is arithmetic, not ideology. Meta's learning phase wants roughly 50 conversion events per ad set per week before delivery stabilises. Split your budget across six ad sets and each one crawls toward that threshold for a month; pool it in one and you cross it in days. The platform's internal 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 manually from the outside.

Two settings deserve attention at this stage. First, conversion objective: optimise for purchase (or your true bottom-funnel event) even in testing — optimising for clicks finds people who click, and click-lovers systematically don't buy. Second, placements: leave them automatic, with one exception — if this is a pure creative test, excluding Audience Network on Meta removes the junk-traffic placements that can flatter a weak video's numbers. Everything else stays default; every manual override during testing is another variable polluting the read.

Set a daily budget of roughly 2–3× your target CPA for the ad set. Anything lower and you'll wait a week for meaningful data. Let the test run for 48–72 hours without touching it. Resist the urge to pause 'losers' at hour six — the learning phase needs room to breathe.

Work the math backwards from the decision you need to make. To confidently kill a creative you want at least 1,000–2,000 impressions on it; to confidently scale one you want several conversions attributed to it. At a $30 target CPA and a $75 daily budget, 72 hours gives you $225 of spend and roughly 5–8 conversions across the batch — enough to rank the six ads, not enough to fine-read them. That's fine: the test's job is to find the top one or two, and precision beyond that is bought during scaling, not testing.

The no-touch rule exists because every edit resets learning. Pausing an ad, changing a budget more than about 20%, or editing copy sends the ad set back into the learning phase, and the two days of data you'd accumulated stop predicting the two days ahead. Schedule the campaign to start at midnight in your account's timezone so day-one data covers a full day, put a calendar block on your own Ads Manager access if you have to, and spend the 72 hours preparing the next batch instead of watching the current one.

Judge creatives in a funnel: hook rate (3-second views ÷ impressions, aim for 30%+), hold rate (15-second views ÷ 3-second views), CTR (aim for 1%+ cold), then CPA/ROAS. A video with a great hook rate but poor CTR needs a new CTA, not a new hook. A video with a poor hook rate is dead — replace the first 2 seconds and relaunch.

The funnel order matters because each metric diagnoses a different two seconds of the video. Hook rate is purely the first moments: the visual, the opening line, the caption. Hold rate is the script's middle — if people stop the scroll but bail at second eight, your problem-agitation beat isn't landing or the pacing sags. CTR is the close: the offer and call to action. Reading CPA first, the way most beginners do, tells you a video failed but never tells you where, which means your next iteration is a guess.

Beware the two classic false reads. A high CTR with terrible conversion rate usually means the ad overpromises relative to the landing page — the fix is on the page or in the claim, and killing the creative wastes a proven attention-getter. And an outlier ROAS on tiny spend (one $200 purchase on $12 of spend) is noise, not signal; don't crown a winner on fewer than three or four conversions. When two creatives look genuinely close after 72 hours, let them run another day rather than picking on gut feel — ties usually break themselves with one more day of data.

After 72 hours: kill anything with spend > 1.5× target CPA and no conversions. For creatives with a strong hook but weak close, generate iterations in Klip Kanvas (same hook, new body/CTA — takes minutes). Move proven winners into a scaling campaign and raise budget 20–30% every 48 hours rather than doubling overnight, which resets learning.

Run the sort as three buckets, in this order. Bucket one, clear kills: poor hook rate or spend past the kill line with nothing to show — pause them and harvest the lesson (which hook style, which avatar underperformed). Bucket two, fixable middles: strong top-of-funnel metrics with a weak close — these go back into Klip Kanvas for surgical iteration, keeping the proven two seconds and regenerating what comes after. Bucket three, winners: hitting or beating target CPA with multiple conversions — these graduate to a separate scaling campaign, leaving the test campaign as your permanent laboratory.

Scale in a duplicate, not in place. Duplicating the winning ad into a fresh scaling campaign preserves the test campaign's structure for the next batch and lets you apply Advantage+ or cost caps to the scaler without disturbing the lab. The 20–30% budget staircase every 48 hours feels painfully slow when a winner is printing, but the alternative is real: overnight doubling routinely sends a 4x ROAS ad back into learning and returns it as a 1.8x ad. Vertical scaling has a ceiling anyway — past it, you scale horizontally by feeding the winner's formula back into step one as next week's batch.

Finally, build a small retargeting ad set for video viewers (50%+) and site visitors, using the static ads from your batch plus a testimonial-style UGC video. Retargeting audiences already know the product — here, offer-led statics ('20% off ends Sunday') convert cheaper than another cold-style video.

Budget it small and deliberate: 10–15% of total spend is plenty, because the audience is only as big as the traffic your cold campaign generates. Set the retargeting window tight — 7 to 14 days for most products — since a visitor from three weeks ago has usually either bought elsewhere or lost interest, and paying to re-reach them drags the ad set's efficiency down. Exclude recent purchasers unless you're deliberately running a cross-sell, or your 'best-performing' retargeting ad will secretly be billing you for customers who had already checked out.

Vary the message by warmth, not just the format. Video viewers who never clicked need the missing information — the testimonial video answering 'does it actually work?' does that job. Site visitors who browsed but didn't buy need urgency or reassurance, which is where the offer-led statics with a deadline earn their keep. Cart abandoners respond best to friction removal: free shipping, easy returns, a simple 'still thinking it over?'. Three tiny audiences with matched messages reliably out-convert one big retargeting pool served the same discount.