Testing vs Scaling Budget Split: The 70/20/10 That Isn't Dogma
70/20/10 gets repeated as gospel, but the right testing vs scaling split depends on your winner inventory and fatigue rate, not a fixed percentage. Here's the framework instead.

One rule of thumb that circulates among media buyers is to put roughly 70% of budget on proven winners, 20% on close variants of ads that are already working, and 10% on genuinely new creative or offer tests. This article cites no platform guideline or study for that 70/20/10 split, so it treats it, and every other number below that has no named source, as an illustrative assumption to adjust, not a benchmark or a rule. Here a "winner" means a creative that has met your own CPA or profit target in your own campaign data, not an ad you have merely seen running. The right split depends on how many winners you currently have, how fast your best creative is fatiguing, and how much runway you have to survive a bad testing week.
Why the ratio isn't the point#
The 70/20/10 framework (or 80/15/5, or 60/30/10, depending on who you ask) gets repeated because it is easy to remember and gives a buyer with no other information a place to start. But it answers the wrong question. The actual decision buyers need to make isn't "what percentage should I test with," it's "how much can I afford to lose this week without threatening the campaign, and how many creative concepts do I need running concurrently to keep a pipeline of winners coming."
Those two questions produce very different splits depending on account size. Take a buyer spending $500 a day in total who tests with 10% of it: that is $50 a day to find a new winner. A buyer at $10,000 a day has $1,000 a day for testing. For illustration only, assume a readable creative test needs about $200 to $250 a day (an assumption for this example, not a typical cost: the real figure depends on your CPA, geo and vertical). On that assumption the first buyer cannot fund even one test, and the second can run four or five in parallel. The percentage is identical; the practical testing capacity is not.
The real inputs to your split#
Winner inventory. If you currently have one working ad and it's showing early fatigue signs, you need to be testing more aggressively than 10%, because you're one bad week from having zero winners. If you have five stable winners across different angles, you can afford to lean harder into scaling and test more conservatively.
Fatigue rate. Native creative fatigue timelines vary by vertical, network and audience size, and the usual explanation is that performance degrades as frequency builds and the same audience segment sees the ad repeatedly. Track how many days your typical winner holds before decay starts, and size your testing budget to have a replacement ready before that point, not after. See creative fatigue for the underlying mechanic.
Account runway. How many bad weeks can the business absorb before spend has to shrink? Thinner runway argues for a smaller, more conservative testing allocation even if it slows how fast you find new winners; the tradeoff is real and there's no formula that removes it.
Testing cost per learning. What a real answer costs depends mainly on your CPA, so work it out from your own numbers. With hypothetical figures, five conversions at a $30 CPA take about $150 of spend, while five conversions at a $90 CPA take about $450. Networks publish their own guidance, and it differs. MGID's help center recommends a test budget of roughly 5,000 to 10,000 clicks and works an example of $600 for 5,000 clicks at a 12 cent CPC. Revcontent's help center suggests letting creatives run for a week, or until you have spent at least $300 to $500, before analyzing the data. Realize (Taboola) states: "The recommended daily budget per campaign is 10 times your CPA goal, or at least $50 USD." Each is that network's own recommendation, not a ranking of which network is cheaper to test on. Budget your test allocation to the cost of a real answer, not an arbitrary dollar figure.
A practical framework instead of a fixed ratio#
- Set a testing floor, not a percentage. Decide the minimum weekly dollar amount needed to run at least two real creative or angle tests to a statistically useful sample. That's your floor regardless of total spend.
- Scale the winner allocation to fill the rest, but cap how much goes into any single creative or publisher, since publisher-level performance data can show that even a strong campaign has weak pockets worth trimming rather than scaling everything proportionally.
- Reserve a middle tier for iteration, not brand-new concepts: variations on a winning hook, a new image on a proven headline, a different CTA. Iterations are usually cheaper to produce than fully new concepts, and they may hit more often, but check that against your own test log instead of assuming it. This middle tier is the role the "20%" plays in 70/20/10.
- Revisit weekly, not monthly (a suggested cadence, not a researched interval). The right split shifts as winners fatigue and new tests either succeed or fail, and a fixed monthly allocation can go stale quickly if your creative decays within weeks.
Where testing budget should actually go#
Not every network or geo deserves equal testing spend. Cheaper geos and networks can let you get a read faster and cheaper, so some buyers test a new angle in a Tier-2 geo on MGID or Revcontent before committing it to a Tier-1 Taboola campaign, because the cost of a wrong answer is lower. We have no data on how common that pattern is, and costs depend on your CPA and on competition in each geo and vertical, so compare real costs before choosing where to test. An angle that works in a cheap geo may not carry over, because the audience, language, offer economics and publisher mix differ, so treat a cheap-geo result as a reason to run a second test, not as proof. Before spending test budget on a concept from scratch, looking at the creatives competitors run in your vertical can suggest angles worth testing. An ad that has been seen running does not show that it is profitable, so use it to choose what to test, not to skip testing.
A worked example across account sizes#
Consider two accounts, both nominally running a 70/20/10 split, with hypothetical numbers throughout. Account A spends $400/day, so its 10% testing allocation is $40/day. Suppose its target CPA is $50. Then $40 a day buys under one conversion a day, about 5.6 in a week for all tests combined ($280 of weekly spend divided by a $50 CPA), which is too few to compare two concepts against each other. It is also below the $50 minimum that Revcontent sets for a campaign's daily budget and that MGID sets for a campaign's overall or daily budget limit (MGID separately asks for a $100 minimum deposit), so it could not fund a separate test campaign on either network. Account B spends $8,000/day, so its 10% is $800/day. At the same $50 CPA that buys about 16 conversions a day. Split across four tests, each gets about $200 a day and about 4 conversions a day, or about 28 in a week. Whether that is a trustworthy read depends on how large a difference between concepts you need to detect.
The fix for Account A usually isn't "test less," it's restructuring the ratio so the testing floor is a fixed dollar amount rather than a percentage. Say the floor is $150 a day (a hypothetical figure). That is 37.5% of Account A's $400 daily spend, not 10%, so the extra $110 a day has to come from the other allocations. If all of it came from the winner allocation, that allocation would fall from $280 to $170 a day under the 70/20/10 starting point. Whether a small account can afford that is a runway decision, and no formula settles it. The alternative is shrinking test quality below the point where it produces usable data. A test that can't reach a real sample isn't cheaper; it spends money without producing a usable answer.
Signals that your split needs to change#
A few example triggers worth watching for, rather than waiting for a scheduled review. The thresholds are illustrative; set your own from your account's history:
- Your best creative's performance has dropped two weeks running. That can be an early fatigue signal (rule out seasonality, bid or budget changes and tracking problems first), and if it is, your testing pipeline needs a candidate ready, not a start from zero once the winner fully decays.
- Your test hit rate has been unusually high recently. If three of your last five tests turned into viable winners, the angle pool you're pulling from may be unusually productive, but five tests is a small sample and luck can explain a streak. Treat it as a reason to test somewhat more for a while and see whether the rate holds, not as proof that the vertical has changed.
- You've had zero new winners in a month despite consistent test spend. This is a signal to look at where your test concepts are coming from, not necessarily to change the ratio; more testing dollars against weak concepts just burns budget faster.
Common mistakes with the split#
- Treating the ratio as fixed regardless of results. If testing is producing winners at a high hit rate, that's a signal to lean in further, not to mechanically cap it at 10%.
- Testing too many variables at once. A test that changes headline, image and landing page simultaneously doesn't tell you which change mattered. Isolate variables where budget allows.
- Killing tests too early. Underfunded tests that get judged on day one or two of spend are usually judged on noise, not signal. Size the test budget to the sample you actually need before deciding.
- Never revisiting "scaling" spend. Money allocated to winners isn't set-and-forget; it still needs ongoing publisher-level and device/geo bid discipline to keep a winner winning.
How to know if your split is working#
Track a simple ratio over a rolling four-week window (a suggested window, not a researched one): dollars spent testing versus new winners produced. If that ratio is getting worse (more testing dollars per winner found), the cause may not be the split percentage; it can be the quality of the concepts going into testing. Better inputs, such as angles you have seen competitors run and your own past results, are worth checking before you reshuffle the budget ratio. A competitor's ad shows that an angle was in use, not that it paid, so test it before you scale it.
The 70/20/10 is one possible place to start on day one of a new account, as an illustrative assumption rather than a researched benchmark. By week four, your own fatigue rate, winner count and test hit rate should be driving the number, not a rule you read in a blog post.







