Most agencies won’t tell you this: the ROAS leaderboard in Ads Manager ranks the past, not the opportunity. It tells you what $5,000 already bought. Scaling asks what the next $1,000 will buy. Those answers can point at different ads.
I used to tell clients to push budget to the winner. I was wrong. An ad can earn a high average because Meta found its cheapest buyers first: people already close to purchasing, warm retargeting pools, placements where the creative fits. More budget cannot be assumed to buy that same mix again. It may reach colder people at higher prices. Delivery changes, then returns change. If you scale by average alone, you can reward the ad that has already used up its best opportunity.
Start with the number the leaderboard cannot show
Average attributed ROAS is simple division: $400 of Meta-credited revenue on $100 of spend is 4.0x. It describes what happened under one attribution setting, inside one window, for one audience mix. It does not quote a price for more growth.
Free tool: the ROAS Calculator works out your return on ad spend, your break-even ROAS and the budget a revenue goal takes.
Marginal ROAS looks at the next chunk. If raising spend from $100 to $200 lifts credited revenue from $400 to $600, the extra $100 returned $200: 2.0x marginal ROAS. The combined average still looks respectable at 3.0x. Both calculations are correct. They answer different questions.
Incremental return asks something stricter: how much revenue would not have happened without the advertising? That is why Meta lift tests compare an exposed group with a holdout and report iROAS rather than platform ROAS. An attributed purchase can be real without being caused by the ad that received credit.
Now add the business constraint. At 40% margin after refunds and fulfillment, break-even ROAS is 2.5x: one divided by 0.40. The 2.0x extra chunk in our example fails that line, even while the account average remains above it. This tripped me up early. I treated a falling average as a warning and a high average as permission. Neither tells you whether the proposed increase pays. Compare the return on the added spend with the break-even line, then ask how much of that return the ad actually caused.

Why more budget changes what an ad buys
At $100 a day, Meta can favor the cheapest conversions available to an ad: past visitors, engaged viewers, the part of a broad audience most likely to respond. Double the budget and delivery has to find somewhere else to spend it. That can mean colder people, different placements, or higher prices in the same auction. It is one reason your creative is your targeting now: change how widely an ad is delivered and you change who sees it. The ad may not perform the same job at twice the spend.
Scale can also change how often people see it. Frequency climbs, the placement mix shifts, and a creative that worked in its best slot has to earn attention in the next one. In one 2.9M-ad dataset, the median Meta ad died within 3 days, three of four were killed within 8 days, and only 5% lasted past 38 days. The cited fatigue pattern was frequency above 2.5 to 3.0, CTR down at least 15%, and CPM and CPA up at least 10%. Those are warning signs, not a timer that tells you when every ad expires.
This is why I want to know what an increase will change before I approve it. If the added budget mostly buys repeated exposure to the same small pool, the attractive historical average is little comfort. Fresh spend needs a plausible source of fresh response.
Watch the winner change when you price the next $1,000
The following numbers are illustrative arithmetic, not observed campaign results. Say you sell a $100 product at 40% margin after refunds and fulfillment, putting break-even ROAS at 2.5x. Two ads each spent $2,000 last week. Ad A drove $10,000 in Meta-credited revenue, a 5.0x average. Ad B drove $7,000, a 3.5x average. The leaderboard points at Ad A. Now estimate what an additional $1,000 would buy from each.
| Ad A: retargeting-led winner | Ad B: broad prospecting runner-up | |
|---|---|---|
| Spend so far | $2,000 | $2,000 |
| Credited revenue so far | $10,000 (5.0x average) | $7,000 (3.5x average) |
| Next $1,000 buys | $1,500 (1.5x marginal) | $2,800 (2.8x marginal) |
| Verdict against 2.5x break-even | Below break-even | Above break-even |
In this example, Ad A’s first $2,000 reached warm buyers cheaply. Its next $1,000 reaches a less responsive mix and returns 1.5x. Ad B started with a broader pool; its next layer returns 2.8x. We would need evidence before assuming a real account follows either curve. But the decision rule is already clear: if those are the available next increments, putting money into Ad A sacrifices margin while Ad B clears the line.
The response to added spend decides the scale bet, not the average on spent money. Here, I would cap Ad A and scale Ad B, despite Ad A’s prettier screenshot.
Check whether the reported sale needed the ad
Retargeting is where the leaderboard can mislead most politely. An ad reaches someone who viewed a product or nearly bought it, then receives credit when that person returns. The report records the purchase. It cannot, by itself, tell you whether the purchase would have happened anyway.
In one agency review, a retention ad showed 6.89x standard ROAS but 0.81x incremental ROAS on new audiences. A supposed loser showed 3.81x standard and 2.15x incremental. The agency’s approach is a useful question for any manager: can you show incremental attribution and audience-segment breakdowns for scale and kill decisions, even if day-to-day optimization stays on 7-day click and 1-day view for volume? A reported purchase is not automatically an incremental one. Your Meta agency can raise ROAS while new-customer growth stalls if it keeps harvesting the warm pool.
The same distinction appears in an automation comparison. Across 640 incrementality experiments, Advantage+ reported 2.4% higher ROAS in Ads Manager, yet 58% of brands saw higher incremental ROAS on manual campaigns. Manual campaigns showed 32% lift against 17% for Advantage+ in the post-treatment window. A platform-credit advantage, in other words, did not consistently mean a causal advantage. If a manager feeds whichever campaign Ads Manager crowns, they may keep rewarding demand capture while starving the work that brings in buyers who were not already on their way.

Build evidence without pretending it is certainty
You can observe how returns change as spend changes. You cannot read a dependable response curve off one week’s ROAS. Step the budget in 20 to 30% moves, hold creative steady, and watch credited revenue per added dollar. It is the same logic I use when I price the next lead instead of the average.
Consider a textbook illustration: $80k of spend drives $280k of revenue, or 3.5x overall. Split it into increments and the picture changes. The first $30k returned $180k at 6.0x; the next $30k returned $75k at 2.5x; the final $20k returned $25k at 1.3x. The useful question is not whether 3.5x sounds healthy. It is whether that final chunk belonged there. The textbook answer is to reallocate until marginal returns equalize across channels. The catch: fitting that curve needs four or five distinct spend levels and clean attribution. Otherwise, the calculation can faithfully reproduce errors in platform credit.
A lift test answers a different question better. Hold out a matched audience, expose the rest, and measure the revenue difference against spend. That gives you causal evidence for the tested conditions, not a guaranteed quote for every later budget increase. It also takes volume. User-level studies typically need 100 or more conversions, run two to four weeks, and often require $50k or more in test spend to separate signal from noise.
With 15 purchases a week, you do not have a precise marginal-return forecast. You have a guess wearing a decimal point. Use budget steps as directional evidence and lift to test causation where volume allows. If neither gives you a confident answer, make a smaller move and set a rollback rule before making it.
Make the person moving budget show their work
Before a manager raises spend, I want a scale plan in writing. It has three parts:
- The hypothesis: Which audience layer will the next dollar reach, and why should that layer respond?
- The money line: What marginal ROAS clears the margin after refunds and fulfillment? In our running example, it is 2.5x.
- The exit: What observation triggers a rollback? For example, roll back an added $1,000 if three-day credited revenue per added dollar stays under 2.0x.
That exit is a decision rule, not a claim that three days reveals the full causal effect. It keeps a weak step from running indefinitely while you examine the result. A manager who cannot write down the hypothesis, threshold, and exit is not scaling. They are hoping with your money.
I also want two kinds of reporting kept apart. One table shows attributed performance by ad, audience segment, and setting; it helps diagnose delivery. A separate note records the causal evidence available: budget-step reads, holdouts, lift results, and the limits imposed by volume. Mixing those into one confident-looking ROAS figure does not make the decision clearer.
This is where I judge providers, too. A percentage-of-spend fee gives the manager more revenue when your budget grows. I prefer the flat model I work inside at groas, where Meta ads run for one $1,499 monthly fee whatever you spend. The incentive is to cut a losing increment rather than keep it running to protect a larger fee. Daily bid fiddling is not a strategy. A defensible next dollar is.
Put the model to work on a new account
Say you spend $20k a month. Retargeting takes $4,000 and reports 6.0x ROAS. Broad prospecting takes $16,000 and reports 3.2x. Your manager proposes shifting $5,000 from broad into retargeting because the leaderboard says retargeting wins. I would not approve that move from those averages alone.
First, set the money line. At 40% margin after refunds and fulfillment, the proposed shift needs to return more than 2.5x at the margin. Next, inspect what the added retargeting budget could buy. Is the audience large enough, or are frequency, CTR, and CPM already signaling repetition? Check broad the same way rather than assuming its next buyer costs what its average buyer cost. Then separate credit from cause: how much of retargeting’s reported revenue depends on reaching people who were already likely to return, and what evidence do you have either way? These are questions to test, not conclusions hidden in the 6.0x.
Suppose the evidence supports an illustrative forecast of 1.4x on the next $5,000 in retargeting and roughly 2.9x on that spend in broad. The decision flips. The retargeting increase misses the 2.5x money line; broad clears it. I would cap the 6.0x ad, keep that money in broad, and plan fresh creative to help hold fatigue down. I would write down the forecast and rollback rule before changing anything, because a forecast is still allowed to be wrong.
That is also how I judge who should run the account. I want someone who states the audience hypothesis, the 2.5x threshold, and the exit before touching budgets; logs each change with its reasoning; and gets paid the same whether spend rises or falls. That is why I work inside the groas setup: continuous execution with a named strategist on a flat fee, where killing a losing increment is the job, not a pay cut. Scale the defensible next dollar, not the prettiest past ratio.
If the account runs under fifty purchases a month, skip the lift-test lecture and use smaller steps instead. I would move 20% of budget, wait a full week, compare credited revenue per added dollar with that 2.5x money line, and roll back when the step misses. That comparison is imperfect; low volume cannot buy certainty. It can still buy discipline.

