Concepts or variations: how to split a game ad test budget
Split the money by what each test costs to read, not by how many ads it holds. What Meta said about similar ads, what it never said, and the arithmetic behind a test budget.
For a mobile game, split test money between new concepts and iterations by what each test costs to read, not by ad count. We'd expect a variation to chase a smaller lift, which takes more installs. By our arithmetic, on install rate alone, 7 concepts at +50% and 3 variations at +20% are 70/30 in ads and 30/70 in money.
Meta separates the two and gives no ratio. We found no data behind the ratios in circulation. So this post prices both kinds of test with arithmetic you can rerun on your own CPI, and says where that arithmetic stops.
Creative iteration vs new concepts: what's the difference in a mobile game?#
A new concept changes the reason to play. A variation, which Meta calls iteration, keeps that reason and changes how it's shown. That's our working test for games, and it sits close to 2 common definitions.
Meta's post of 16 December 2025 describes diversification as distinct assets made for different personas or use cases, and iteration as manual A/B testing that refines assets, formats and messages. Its example of iteration is 2 ads with the same visuals and different text CTAs. RocketShip HQ, an app-growth agency, draws the line at the bet: change the claim, the pain or the reason to act and you have a new concept, even if the ad looks similar.
In a game, that sorts like this:
- New concept: "relax with 5 minutes of sorting" against "beat the player above you". Different reason to play, so a different player might stop. Quantic Foundry's motivation model, built on survey data from over 140,000 gamers, is one starting list of reasons.
- Variation: the same "beat the player above you", with the hero swapped for the villain, or a new headline, background or button.
Games make the line harder to hold in one way. Every ad for a game shares its art style, its characters and its logo, and the studio wants it that way. So 2 real concepts can still look alike to a system that groups ads by what they look and feel like. Meta's advice is ads that differ in look, feel, storyline and message. It doesn't say how that applies when the art style can't move. Our suggestion is to change what the style allows: the scene, the action, the character's role and the format.
What Meta said about similar ads, and what it didn't#
Meta said its system groups similar ads. It gave no threshold and no budget advice. Here is what bears on a test budget, all from the December 2025 post:
- Iteration and diversification are separate. Iteration can inform new concepts but shouldn't replace diversification.
- The system groups ads that share key visual and thematic attributes. Ads that look or feel alike count as variations of one creative, so "learnings and delivery optimizations are shared at the creative level".
- Similarity and fatigue are different problems. Fatigue is overexposure and similarity is too little visual diversity, though high similarity can also drive fatigue.
- Instead of a fixed volume of refined ads, Meta suggests a broad portfolio of distinct ones.
The post's performance figures cover Meta's own generative tools and Advantage+. None of them measures a split.
| What you'll read | In Meta's post? | Where it comes from |
|---|---|---|
| Iteration and diversification are separate | Yes | Meta, December 2025 |
| Similar ads are grouped and share learnings and delivery | Yes | Meta, December 2025 |
| Iteration has stopped being worth doing | No. Meta says it can inform new concepts | An overreading |
| Ads above 60% similarity get collapsed or suppressed | No. The post states no number | Agency and vendor blogs |
| Grouped ads share one Entity ID | No. The term isn't there | Industry shorthand |
| Run a set number of ads per ad set | No. The post states no count | Agency and vendor guides |
| A budget split between concepts and variations | No | No one with data that we could find |
The 60% figure has a trail you can follow. Chatterbuzz Media, a performance marketing agency, says scores above 60% collapse similar ads into one Entity ID, and links no Meta document for it. Ads Uploader, a vendor, calls Entity ID industry shorthand and not a term Meta publishes. Brandsearch, another vendor, says Meta has never published the numbers attached to the similarity score, and rewrote its own page to stop repeating them. We searched Meta's help pages on 31 October 2026 and found no score or threshold either. Meta does say it's testing creative insights in Ads Manager that help diagnose similarity, so an account may show more than the public pages do.
Our read for budgets is that a variation doesn't buy a new slot. If it looks like its parent, Meta's system treats the two as one creative, so adding it shouldn't open new audiences. It may still be the better ad, so let a winning variation replace its parent, and count your slots in concepts. Meta doesn't state this about tests. It's our inference. The post is also about Meta alone, so don't assume TikTok, Google or the game networks behave the same way.
There's no benchmark for the split#
We found no dataset, for games or anything else, that measures how a concept-to-variation split changes results. What circulates is opinion, and some of it is useful:
- Segwise, a creative analytics vendor, offers 70/20/10 of creative output: 70% iterating on winners, 20% new concepts in familiar formats, 10% experimental. Elsewhere it suggests roughly 50/50 for established brands. Neither is written for games.
- RocketShip HQ says no fixed ratio is worth copying. Several fresh winners pull the week towards variants, and one tiring winner pulls it towards new concepts.
- Unity, writing for game teams, says high-spend teams generally need new concepts every week or two, with variants of winners filling the gaps. That's a cadence rather than a ratio.
All of these count ads or output, and that's the gap. An ad in a concept test and an ad in a variation test don't cost the same to read.
A variation usually costs more to read than a concept#
A test's price is set by the smallest difference it has to read, and variations usually chase smaller differences than concepts do.
The claim about variations is reasoning, not a measurement. A variation shares its bet with the control, so we'd expect a narrower gap, while a new concept can land far above or far below. We found no public dataset that compares the two in games, so check it against your own history.
The claim about price is arithmetic. By a read, we mean an 80% chance of detecting a difference at 5% significance. We ran the standard two-proportion sample-size calculation, the one in R's power.prop.test, with equal impressions per cell and a two-sided test. Then we checked each result with 200,000 simulated tests. Install rates of 0.5% and 1% per impression gave almost the same install count, so the chart holds for any low install rate.
| +100% | 23 |
|---|---|
| +50% | 78 |
| +25% | 281 |
| +20% | 429 |
| +10% | 1,640 |
Multiply by your CPI to price a cell. At the $2 CPI Unity uses as its example, a +50% read costs about $156 a cell and a +20% read about $859. Each test also needs a control cell of that size. Evan Miller's calculator, a common check, uses a slightly different approximation and comes out about 14% lower at +50%, so expect small gaps if you rerun this.
Unity's example is 50 installs a variant, about $100, and it tells you to derive your own number. That size suits a big swing. We simulated it with 10,000 impressions a cell at a 0.5% install rate, which is about 50 installs in the control, over 200,000 runs per case at the same 5% test. That cell reached significance on a true +50% lift 61% of the time, and on a true +20% lift 16% of the time. The real +20% variation tied or lost on raw installs in 18% of tests. And 2 identical ads showed a gap of 20% or more in 37% of tests.
Segwise's 2026 testing guide says about 50 conversions per variant detect a 20% difference at 95% confidence, and that smaller gaps need far more. It's right about the direction. It names no power, and at 80% power we get about 430 for that gap.
Why a 70/30 split of ads can be 30/70 in money#
Because one +20% read costs about 5.5 times one +50% read, 3 variations can outspend 7 concepts. Take 10 test ads a month, and read the 7 concepts at +50% and the 3 variations at +20%, at a $2 CPI:
| Cells | Installs per cell | Spend | |
|---|---|---|---|
| New concepts | 7 | 78 | $1,092 |
| Variations | 3 | 429 | $2,576 |
By ad count that's 70/30. By money it's 30/70. Add a control cell to each side and it moves to about 27/73. Spend uses unrounded installs (429.3 a variation cell).
| 7 new concepts, read at +50% | 30% |
|---|---|
| 3 variations, read at +20% | 70% |
The reads decide it. Read the variations at +50% too and the money goes back to 70/30, at $1,560 in total. Read them at +10% and it goes to 10/90, at $10,930. That's the decision a ratio hides: how small a lift is worth paying to see.
All of this reads install rate, and Unity warns that a creative can win on IPM and lose on ROAS. A new concept is meant to bring a different player, so it needs a retention read more than a variation does. By the same formula, at an example 15% D7 rate, reading a +50% retention lift takes about 424 installs a cell, about what a +20% install-rate read takes. If only the concepts get that read, the money goes back to about 70/30. If both sides get one (+20% takes about 2,400 installs a cell), it stays near 30/70. So decide which tests need a retention read before you split the money. We haven't priced a ROAS read.
A small lift is only worth reading behind real spend. At a flat CPM, a +20% install rate buys 20% more installs from the same money. A control and a variation at that read cost about $1,720 at a $2 CPI, so the test pays back once about $8,600 of spend runs behind the winner, if the lift is real. Divide by your hit rate: if 1 variation in 5 holds a real lift, that's about $43,000. This treats the test spend as lost, values each extra install at your CPI and ignores the real lifts a test misses. The first makes the figure too high and the last too low, so use it as a rough hurdle with your own inputs.
How to set your split#
Use your account's state for direction, then the arithmetic for the money. The table below is our judgement, and its direction matches RocketShip HQ's above.
| Your account | Where test money goes | Why |
|---|---|---|
| No proven winner: a new game or a new geo | Nearly all on concepts, read at big swings | There's nothing to vary yet |
| One winner carrying spend, and tiring | Concepts first, plus a few variations of it | You need its replacement, and time to find it |
| Several healthy winners | Most on variations, with concepts in every round | Variations explain what won. Concepts find what's next |
- 1Set the test budget from your own account
Decide how much you can spend on tests that may all lose. Meta's ceiling for a creative test, below, is a limit for one tool and doesn't tell you what to spend.
- 2Pick the smallest lift worth reading on each side
Our starting points are +50% for concepts and +20% for variations. Move them, since a smaller read only pays behind bigger spend.
- 3Price the cells
Installs from the chart × your CPI, plus a control cell per test. Add a retention read where the test needs one.
- 4Fit the round to the money
If a +20% read doesn't fit, test a bigger change, or ship the small one on judgement and don't call it a result.
- 5Keep a hit rate for each side
Count wins per side and re-price monthly. Your own history replaces our reasoning about effect size.
On Meta you have 2 tools, and neither prices the test for you. A creative test takes 2 to 7 copies of one ad and suggests at most 20% of an existing budget, to limit the effect on the campaign. It names top performers, and its help page says the results include no confidence level. Jon Loomer notes the original ad stays out of the test, so leave one copy unchanged as your control. An A/B test makes sure nobody sees both versions and reports a winner with a confidence percentage. Meta's best practices ask for at least 7 days, and for a budget that produces enough results, but give no figure for it. We'd screen with the first and use the second where you need a read. Our hook test guide runs the same method on openings.
This arithmetic leaves other things out. It assumes cells are split at random, as in an A/B test. More cells mean more comparisons, which calls for stricter reads. Stopping at the first good-looking result inflates false winners. And a test built at 80% power for a given lift misses that lift 20% of the time.
A variation only gives a clean read if one layer changes#
If 2 things changed, the result can't say which one mattered. Meta's best practices and Unity's guide both ask for one change at a time. So build every branch from one control, and lock each layer except the one under test.
- background
- props
- character
- logo
- headline · copy
Clash of ClansStrategy · FreeBUILD YOURVILLAGE.- Build & defend
- Raid for loot
- Join a clan
- button
- the adClash of ClansStrategy · FreeBUILD YOURVILLAGE.
- Build & defend
- Raid for loot
- Join a clan
Play free→
| Branch | Layer that changes | What the result answers |
|---|---|---|
| Control | None | The baseline |
| A | Character: the hero becomes the villain | Does this character pull more installs? |
| B | Headline: a progress claim becomes a challenge | Does this line beat the old one? |
| C | Background: day becomes night | Does the setting matter? |
| D | Button: a new label | Does the label matter? |
Those 5 cells fit inside Meta's 2 to 7 copies. At a +20% read and a $2 CPI they cost about $4,300, which is a lot to spend on a button label, so run the branches you most need answered.
The rest has to hold still. Regenerate a flat picture to change one word and the character and background can shift with it, so the test now measures several changes at once. In a layered ad, the layers you didn't touch stay as they were.
That's the job we built Vidscape for. Each ad in a round comes in layers, such as character, headline, background and button. A round is 1 to 14 ads, each a different angle, in 1:1, 9:16 or carousel, and characters are redrawn with your game's art as the reference. That feeds the concept side, though whether 2 angles count as distinct is Meta's call, and Meta doesn't say. The variation side is done by hand: open the ad in the editor, change the one layer, export, undo, change the next. There's no one-click version maker, and Vidscape doesn't read campaign results, so the read stays in your ad manager or attribution tool. More on Vidscape for growth teams.
Frequently asked questions#
What's a good split between new concepts and variations for a mobile game?
We found no public benchmark, and Meta gives no ratio. Work from your account. With no winner, nearly all test money goes to concepts. With one tiring winner, lead with concepts and extend it with a few variations. With several healthy winners, spend more on variations and keep concepts in every round. Then price each side by the lift it must read, because the money split can differ sharply from the ad split.
Does Meta penalise similar ads?
Meta doesn't use that word. Its December 2025 post says ads that look or feel alike are treated as variations of one creative, with learnings and delivery shared across them, and that high similarity can also drive fatigue. It states no similarity score or threshold. The 60% figure on agency and vendor blogs isn't in that post, and we didn't find it in Meta's help pages.
How many installs does a creative test need?
It depends on the smallest lift you want to read. At 5% significance and 80% power with equal impressions per cell, our arithmetic gives about 23 installs in the control cell for a doubling, 78 for +50%, 430 for +20% and 1,640 for +10%. That reads install rate only, and a retention read takes more: about 424 installs a cell for +50% at an example 15% D7 rate. About 50 installs a cell suits big swings, not small lifts.
Can a small studio test variations at all?
Yes, but only big changes. By our arithmetic, at a $2 CPI (Unity's example), reading a +20% lift costs about $1,720 for a control and one variation, before any retention read. If that doesn't fit, test changes large enough to show at +50%, such as a different opening, or ship small changes on judgement and don't treat them as test results. Put the rest into new concepts, where we'd expect wider gaps.
- Demystifying creative diversification — Meta for Business · Dec 2025
- Set up a creative test in Meta Ads Manager — Meta Business Help Center · read Oct 2026
- About A/B testing — Meta Business Help Center · read Oct 2026
- How winning campaigns are determined in A/B tests — Meta Business Help Center · read Oct 2026
- Best practices for A/B testing — Meta Business Help Center · read Oct 2026
- What is creative testing? A guide for mobile game creatives — Unity · Oct 2026
- What is a creative concept vs a creative variant? — RocketShip HQ · Jun 2026
- How to iterate winning creatives — Segwise · May 2026
- How many ads to produce per month — Segwise · Mar 2026
- Ad creative testing best practices — Segwise · Jun 2026
- Meta's creative testing tool — Jon Loomer Digital · Jan 2026
- Meta Andromeda: AI creative targeting guide — Chatterbuzz Media · 2026
- Meta Andromeda explained — Ads Uploader · Jun 2026
- Meta creative similarity score — Brandsearch · Apr 2026
- Gamer motivation model — Quantic Foundry
- power.prop.test: power calculations for two-sample test for proportions — R documentation
- Sample size calculator — Evan Miller
