Vary the hook first, the argument second and the opening shot third, because those three are the only changes that reliably move a number you are being judged on. Length, CTA wording, captions and music are worth testing after those are settled, and font, colour grade and logo placement almost never repay the effort. Aspect ratio is not a variation at all: four crops of the same ad are one ad in four placements.
Most teams get this order wrong in a specific way. They test the things that are easy to change in an editor, because those are the things they can change in an editor, and the proposition the ad is arguing goes untested for a quarter.
The ranked list, and the metric each change shows up in
| Rank | What you change | Why it moves the number | Diagnostic metric | Cost to try |
|---|---|---|---|---|
| 1 | The hook: first line plus first frame, 2 to 3 seconds | Gates every downstream metric. Nobody sees your argument if they do not get past it | Hook rate, 3-second view rate | New variation, 1 credit |
| 2 | The argument: the claim the ad actually makes | Decides whether the traffic converts, which the hook cannot touch | Conversion rate, cost per result | New variation, 1 credit |
| 3 | Opening shot subject: who or what is on screen at frame one | A face, a hand and a pack shot stop different people | Thumbstop rate | New variation, 1 credit |
| 4 | Proof type: testimonial, demo, or a number on screen | Moves the middle of the ad, where doubt gets resolved | Hold rate at the halfway mark, conversion rate | New variation, 1 credit |
| 5 | Structure: proof before the demo or after it | Reorders when the payoff arrives | Completion rate | New variation, 1 credit |
| 6 | Length: 15 seconds against 30 | Shorter usually completes more, longer usually qualifies harder | Completion rate, cost per result | Free edit |
| 7 | Voiceover: voice, pace, or the language it is written in | A market-native script beats a translated one, audibly | Completion rate, conversion rate by market | New variation, 1 credit |
| 8 | CTA wording and where it lands in the timeline | Small on CTR, occasionally large on lead quality | CTR to site, conversion rate | Free edit |
| 9 | Caption style and burn-in | Feeds play muted, so captions carry the script | Completion rate on sound off | Free edit |
| 10 | Music bed | Real but small, and easy to over-attribute | Completion rate | Free edit |
| Not a variation | Aspect ratio, colour grade, brand font, logo placement | Placement compliance and brand consistency, not creative variety | None worth reading | Free edit |
Why the hook sits at the top, in numbers
An ad's click count is two multipliers, not one:
clicks = impressions × hook rate × click rate among viewers who got past the hook
Take the Meta feed median from our 2026 fatigue benchmark, which is a synthesis of published platform and agency figures rather than our own measurement: 1.62 percent video CTR and a 28 percent median hook rate. Most accounts we see run a Meta feed hook rate somewhere between 15 and 25 percent, so work the example at the bottom of that band. On 100,000 impressions at a 15 percent hook rate and 1.62 percent CTR you get 1,620 clicks, and the implied click rate among the 15,000 who stayed is 1,620 / 15,000, or 10.8 percent.
Now move one thing at a time.
- Hook rate 15 to 25 percent, downstream click rate held at 10.8 percent: 25,000 × 10.8 percent = 2,700 clicks. CTR goes 1.62 to 2.70 percent, a 67 percent lift.
- Hook rate held at 15 percent, downstream click rate improved from 10.8 to 12.5 percent, which is a good win from a rewritten body: 15,000 × 12.5 percent = 1,875 clicks. CTR goes 1.62 to 1.88 percent, a 16 percent lift.
Four times the effect for the same amount of work, which is why the hook is rank one. The reason the argument is rank two rather than rank six is that the hook cannot fix the wrong promise. Cheaper clicks into a proposition nobody wants is just more expensive traffic. Test hooks to fix cost per click, test arguments to fix cost per acquisition, and know which of those two numbers is currently hurting you before you pick a dimension.
The bottom half of the table is not worthless, it is second order. A caption restyle that lifts completion by a couple of points is a real gain on a creative that is already winning. It is a waste of a week on a creative nobody watches past second two.
Aspect ratio is not a variation
Genyad exports 9:16 at 1080x1920, 4:5 at 1080x1350, 1:1 at 1080x1080 and 16:9 at 1920x1080, and all four are the same ad. A viewer who sees your hook in a 4:5 crop and then in a 9:16 crop has seen your hook twice, which is what the benchmark's finding that CTR drops 45 percent after a fourth exposure to the same creative is actually counting.
This matters because of how rotation gets reported. Five ads exported to four ratios shows up in a spreadsheet as 20 assets, gets described in a stand-up as 20 variations in rotation, and behaves in the auction like five. The ad variation glossary entry draws the line properly. Crops are a QA job: check the safe zones, confirm the captions are not under the interface furniture, ship all four. Re-exporting costs nothing in credits, so there is no reason to ration ratios and no reason to count them.
One change at a time, because attribution is arithmetic
The rule sounds like methodological fussiness until you cost it. Take a 5,000-impression read per cell as a planning number and a €12 CPM, which puts one read at 5,000 × 12 / 1,000 = €60.
| Test design | Cells to fill | Impressions needed | Cost at a €12 CPM | What you learn |
|---|---|---|---|---|
| One dimension, 4 levels | 4 | 20,000 | €240 | Which of 4 hooks wins, cleanly |
| Two dimensions, 4 levels each, full factorial | 16 | 80,000 | €960 | Hook and length, plus their interaction |
| Three dimensions, 4 levels each, full factorial | 64 | 320,000 | €3,840 | Rarely worth it outside a large account |
| Three dimensions, changed all at once in one ad | 1 | 5,000 | €60 | One fact and three hypotheses |
The last row is what most teams actually do, and it is the row with the worst return despite the lowest price. You get a winner you cannot decompose, so next month you rebuild the same ad and guess which part mattered. Sequential single-dimension tests cost 12 reads and €720 to cover the same three dimensions, and every one of the three answers is attributable.
The practical version: hold the hook, body and CTA structure fixed and move exactly one slot. Ten hooks against one body and one CTA is a hook test. Ten hooks against ten different bodies is a batch of creative, which is a legitimate thing to ship, it is just not a test and should not be written up as one.
Which changes are worth a credit and which are free
Anything that regenerates the script or the voiceover is a new variation and costs 1 credit. Anything you can do to an existing variation on the timeline is free, and so is re-exporting and uploading footage. That split lines up almost exactly with the ranked table: ranks 1 to 5 and rank 7 cost a credit, ranks 6, 8, 9 and 10 do not.
Which is a convenient way to sanity-check a testing plan. If your last month of creative work cost you zero credits, you spent the month on ranks 6 through 10. Genyad is our product, so treat this as disclosure: the free plan gives 5 variations with no card, Starter is €29 for 15 credits and Growth is €99 for 65 credits, and credits never expire. The video ad editor is where the free changes get made, including CTA text, caption styling, trims and the four ratio exports.
Two honest limits. Genyad has no predicted performance scores, so it will not rank your ten hooks before you run them, and if a model-scored pre-flight is what you want, AdCreative.ai was listed at $39 per month in August 2026. It has no AI avatars or synthetic presenters, so presenter type is not a dimension you can vary with us. And if the only thing you need is hook and CTA swapping inside one locked brand timeline, modular template rendering is cheaper per render than generating a fresh variation: Sovran was listed at $99 per month in August 2026. Prices move, so check them yourself.
Frequently asked questions
What should I vary first in an ad variation?
The hook: the first line and the first frame, which is the first 2 to 3 seconds. It is the only change that multiplies every metric downstream of it, and at a 15 percent hook rate lifting it to 25 percent raises CTR by 67 percent while a strong rewrite of the body raises it by roughly 16 percent. Vary the argument second, because a better hook cannot rescue a promise nobody wants.
Does changing the aspect ratio count as a new ad variation?
No. Exporting one ad at 9:16, 4:5, 1:1 and 16:9 gives you four files and one creative, and a viewer who sees two of them has seen the same ad twice. Treat crops as placement coverage and a safe-zone QA check, then count only distinct hooks, arguments and shot selections towards your rotation.
Should I change one thing at a time or ship batches?
Both, for different jobs. A test changes one dimension and holds the rest, so the result is attributable: four hooks against a fixed body and CTA costs four reads. A batch changes everything at once to keep a rotation stocked, which is necessary work but produces a winner you cannot decompose, so do not report it as a test result.
Is music worth testing in video ads?
Rarely, and never before the hook is settled. A music swap shows up in completion rate and occasionally moves it a point or two, which is a real gain on an ad that already holds attention and irrelevant on one that loses 80 percent of viewers in the first two seconds. It is also a free edit, so it costs nothing but the week you spent on it.