Version creative along one axis at a time and hold everything else constant. Change the hook or change the offer or change the format, never two at once, because a variant that differs in three ways teaches you nothing when it wins. The discipline sounds obvious and is broken constantly, usually because generating variants is now so cheap that people produce forty and can no longer say what any of them proved.
Setup
You need the key visual or the master cut, the list of placements with their format requirements, and a decision about what you are actually trying to learn. That last one determines the axis. Versioning without a learning goal produces volume, not information.
What to hold constant
The idea. Every variant is the same campaign. If two variants would read as different campaigns to a viewer who saw both, you have split your budget across two campaigns.
The brand furniture. Logo treatment, typography, palette, end card. These are never the test variable in a performance campaign, because a win driven by a bigger logo is not a learning you can use.
The offer, unless the offer is the axis. Changing price framing and hook at once is the single most common way a test becomes uninterpretable.
The axes worth testing, in order
1. Hook. The first two seconds carry more variance than everything else combined. Five hooks against one body is the highest-value test in short-form video, and the cheapest to produce.
2. Format and aspect ratio. Not really a test, more a requirement, but versioning here is where campaigns silently lose quality when a 16:9 master gets machine-cropped to 9:16 and the subject drifts out of frame.
3. Proof. Demonstration versus testimonial versus result-only.
4. Offer framing. Free shipping versus percentage off versus bundle.
5. Length. The 30 and its 15 and 6 second cutdowns, which should be edited rather than trimmed. A cutdown that is just the first six seconds is not a six-second ad.
How many variants
Three to five per axis. Below three there is no comparison; above five the traffic per variant gets thin enough that the results stop meaning anything unless the spend is large.
The math that matters is per-variant volume, not variant count. Forty variants on a small budget produce forty statistically meaningless results, which feels like rigor and is not. Creative fatigue is a real reason to refresh volume; it is not a reason to split a small budget forty ways.
Naming, which is not admin
A convention that survives contact with a media team: campaign_axis_variant_format_date. For example refill_hook_02_9x16_0828. Three properties make it work: the axis is in the name so results are sortable by what was tested, the format is explicit so nobody uploads the wrong crop, and the date lets you find the version six weeks later when someone asks which one ran.
Every file that leaves the team has this name. Files named final_v3_REALLY_final cost hours later.
What versioning costs now
Using canon 8frame prices, a credit being $0.01 at pack rate: five hook variants generated as 8-second clips on Veo 3.1 Lite with audio is 270 credits, about $2.70. The same five on Kling v3 Standard at 5 seconds with audio is 430 credits. Five image variants on Seedream 5 is 30 credits.
Cheap enough that the constraint is no longer production. It is interpretation: the discipline of one axis at a time, and the willingness to kill variants that won by accident.
When to stop versioning
When the variants stop differing in ways a viewer would notice. Twelve versions of the same hook with different word order is not a test, it is a way of avoiding the harder work of a second idea. See AI ad variant testing workflow for the testing structure itself.
FAQ
How many ad variants should a campaign have? Three to five per tested axis, sized so each variant gets enough traffic to produce a readable result.
Should variants be generated or edited from a master? Format versions get edited from the master so the idea stays intact. Hook variants get generated, because they are genuinely different content.
What is the most common versioning mistake? Changing two things at once, then attributing the win to the one you preferred.
Versions are easiest to keep straight when they sit side by side on one surface instead of in a folder tree. The 8frame canvas is free and unlimited, and generation is paid from $19/month.