A typical active Meta campaign needs 8 to 20 live variations in rotation, refreshed on a schedule set by how fast your vertical decays rather than by how the ads feel. Variations only count if they differ in hook, structure and framing, because two edits of the same timeline with different colour grading read as one ad to the delivery system and to the person scrolling past. The operational question is not whether to make more, it is how to produce a readable set on a repeating cadence without the production cost growing with the count.
How many variations should be live at once?
Eight to 20 live variations per active campaign is the range we work to. The lower end is for a single product with one clear audience, the upper end is for multi-market or multi-audience campaigns where each segment needs its own angle.
Monthly throughput matters more than the live count. Our 2026 fatigue benchmark, a synthesis of published platform and agency figures, finds that brands shipping 15 to 50 creative variants a month see 3 to 5 times longer campaign lifespan than teams refreshing quarterly, and the report's own conclusion is that throughput rather than talent is the bottleneck: the number of unique concepts a team ships per month predicts campaign longevity better than the quality of any single ad.
That is uncomfortable if your process is one shoot per quarter. To size a set against your own spend and refresh rate, the ad variation calculator does the arithmetic.
There is a real floor. If your budget cannot get eight ads out of learning, eight is worse than three: you get eight underpowered reads and no decision. Small accounts should run fewer variations and refresh them more often.
What does distinct mean to Meta's optimiser?
Delivery treats each ad as a separate thing to be tried, priced and either fed or starved, so the question is not whether two variations are different files but whether they behave differently in the auction. Ours ranks like this:
| Change | Reads as a new ad | Why |
|---|---|---|
| New hook, first three seconds | Yes, strongly | Decides hook rate, which decides everything downstream |
| Different structure or shot order | Yes | Changes where the offer lands and who is still watching |
| Different angle or claim | Yes | Different people self-select into it |
| Different framing and ratio per placement | Yes | Composition, not crop, changes what is readable |
| New voiceover or language | Yes for language, partly for voice | Native-language scripts pull different audiences |
| Caption style and pacing | Marginal | Matters for muted playback, rarely changes who clicks |
| Colour grade, typeface, music swap | No | Same ad wearing a hat |
| Same edit trimmed to a different length | Barely | Inherits the original's structure and offer placement |
The definition we use for an ad variation is a version built to test a specific difference, not a re-export. The practical test: if you cannot say in one sentence what this variation is asking, it is not a variation, it is a duplicate that will split your data.
This is also why re-cutting one hero timeline hits a ceiling fast. Every cut inherits the same shot list, the same offer position and the same pacing, so after about four versions you are producing colour grades in the table above.
Should you ship the whole set at once or drip-feed it?
Ship in cohorts. Our default is five to eight variations launched together, run for the same window, then judged against each other, with the losers cut and replaced by the next cohort. Simultaneous launch is what makes the comparison mean anything, because ads launched two weeks apart competed in different auctions against different creative, and the later one usually looks better for reasons that have nothing to do with the ad.
Drip-feeding has one honest advantage: it keeps delivery stable. Adding a large cohort to a mature campaign can shake up allocation while the new ads are priced, and if you are running close to your CPA ceiling that is a real cost. So on small budgets, or on a campaign you cannot afford to disturb, add two or three at a time and accept that you are getting a rougher read.
What does not work is the middle path most teams land on: one new ad whenever someone has time. That gives you neither stability nor comparability, and the oldest ad in the account usually keeps the budget because it has the longest history.
Cadence should follow decay speed, not the calendar:
| Vertical | Days to a 40 percent CTR decline | Cohort cadence we would run | Live set size |
|---|---|---|---|
| Food and beverage | 9 | Weekly | 12 to 20 |
| Fashion | 12 to 14 | Weekly to fortnightly | 12 to 20 |
| Beauty and DTC | About 18 | Fortnightly | 8 to 16 |
| Electronics | About 21 | Fortnightly to monthly | 8 to 12 |
| B2B SaaS | About 28 | Monthly | 8 to 12 |
The decay figures come from our benchmark report. The cadence and set-size columns are our working defaults, not findings.
How to read results per creative
Read in this order, and stop at the first metric that explains it.
Hook rate first. It tells you whether the ad failed to start or failed to convince, which have opposite fixes. Median hook rate sits around 28 percent on both Meta Reels and Meta feed, so anything much below that is an opening problem, and rewriting the body of the ad is wasted work.
Then hold, then click. Median video CTR runs about 1.78 percent on Meta Reels and 1.62 percent on Meta feed. An ad with a healthy hook and a weak CTR has a middle problem, usually an offer that arrives too late or a claim the footage does not support.
Then frequency. Weekly frequency around 2.5 on Meta prospecting is where decline begins in the benchmark data, and Meta's internal research puts the CTR drop at 45 percent after a fourth exposure to the same creative. Rising frequency with falling CTR is fatigue, not a bad ad, and the fix is a replacement rather than a rewrite.
CPA last, and only with enough volume behind it. Killing on CPA at low volume is how teams cut ads that were fine, and it is why cohort launches beat rolling ones: you get comparable reads on the upstream metrics well before CPA becomes trustworthy.
One caution on per-asset breakdowns. With Meta's automated creative features on, those numbers tell you what delivery chose, not what a clean test would show, because allocation is not random. Use them to spot which assets never got served, then ask why delivery found them unattractive.
Where the throughput comes from
Disclosure: Genyad is our product, and this is the problem it exists to solve. You upload the footage you already own once, it transcribes and tags every clip, and each variation is generated as its own script, shot selection, voiceover, caption set and export drawn from that library rather than a re-cut of one timeline. That is what makes a weekly cohort of eight possible without eight edit sessions.
Each variation costs one credit, exports in 9:16, 4:5, 1:1 and 16:9, and scripts are written natively in English, German, French, Spanish, Italian or Hindi rather than translated, which matters for multi-market sets. Editing and re-exporting cost nothing. The Meta ad creative generator page shows the flow, and the free plan is five variations without a card.
It does not publish to Meta, so shipping a cohort still means uploading files in Ads Manager yourself. There are no AI avatars, no static banners, no product-feed template rendering and no predicted performance scores: if you wanted a score telling you which of the eight wins before launch, we do not have one.
Frequently asked questions
How many ad variations should I run in one Meta campaign?
Eight to 20 live variations suits most active campaigns, with the lower end for a single product and audience. The constraint is budget: if your spend cannot get eight ads out of learning, run three and refresh them more often. Monthly throughput matters more than the live count, and 15 to 50 new variants a month is where our benchmark report sees campaign lifespan stretch 3 to 5 times.
What counts as a genuinely different ad variation?
A different hook, structure, claim or framing. A different colour grade, typeface or music bed is the same ad, and Meta's delivery treats it that way while your data gets split across duplicates. The test we use is whether you can state in one sentence what the variation is asking.
Should new Meta ads be launched all at once or gradually?
Launch cohorts of five to eight together so they compete in the same auctions and can be compared. Gradual addition keeps delivery stable, which is worth something on small budgets or a campaign near its CPA ceiling, but the reads are rougher. Adding one ad whenever someone has time gives you the worst of both.
How often should Meta ad creative be refreshed?
Follow your vertical's decay speed. Food and beverage hits a 40 percent CTR decline in about nine days, fashion in 12 to 14, beauty and DTC around 18, electronics around 21 and B2B SaaS around 28, according to our benchmark report. That maps to weekly refreshes at the fast end and monthly at the slow end.