MER vs ROAS and which one to steer your ad spend by

Four numbers get called ROAS in the same conversation and they answer four different questions. Platform ROAS is one model's opinion of its own work. MER is the whole business divided by the whole ad bill. aMER isolates acquisition. Incremental ROAS asks what wouldn't have happened without the spend, and as of May 2025 it is the one that stopped being enterprise-only.

What do platform ROAS, MER, aMER and incremental ROAS each measure?

They sit at four different scopes, from one ad account to the whole business. Getting them confused is the most expensive measurement error in ecommerce, because each one is genuinely right about its own question and badly wrong about the others.

Note what MER doesn't require. No pixel, no attribution window, no cooperation from a platform. It's two numbers you already have, and that's the whole appeal — it cannot be gamed by an attribution setting because it never consults one.

Do the ad platforms overstate performance?

Not uniformly, and the direction depends on which campaign type you are asking about. The largest public dataset on this is Haus, which ran 640 Meta incrementality experiments across advertisers averaging $14M a year in Meta spend and published the aggregate in July 2025.

Read the first two together, because separately each supports a different tidy story and neither story is true. Meta as a channel was worth more than its dashboard claimed — the average lift to the primary KPI across those experiments was about 19%, with the strongest single test at 74%. Meta's automated campaigns, meanwhile, credited themselves with more than they earned relative to manual ones, and 58% of brands got a higher incremental ROAS out of manual.

The other reason this matters more each year is price. Meta's own FY2025 results reported impressions up 12% and average price per ad up 9%, on $196.175B of ad revenue. Measurement error is charged at the going rate, and the going rate keeps rising.

Why did brands that split budget evenly between automated and manual do worse?

Because a split budget starves both halves of the conversions each needs to learn. Haus watched a dozen or so brands move from a heavily skewed split to roughly 50/50 between Advantage+ and manual campaigns, and their incremental ROAS fell about 18%. That is a before-and-after within the same brands rather than a comparison between brands, and it is a small sample — but it points the same way as the learning-phase arithmetic. Hedging lost to both commitments.

The likely mechanism has nothing to do with which type is better. Meta needs around 50 optimisation events per ad set per rolling seven days to leave the learning phase, and halving a budget across two structures doubles the number of ad sets that have to clear that bar on half the money. It is a signal problem wearing a strategy costume. The launch guide works through what that means for a starting budget.

Worth being precise about what has changed structurally too. Since February 2025 Advantage+ is not a campaign type you select. It is a state a Sales campaign is in, and turning off budget, audience or placement automation turns it off, while keeping some automated behaviour. So the clean A-versus-B this finding describes is harder to set up now than it was when the experiments ran.

What is MER good for, and where does it break?

MER is the number to steer the business by, and the case for it is that it can't lie about attribution. Total revenue over total paid spend has no window, no model and no incentive. If MER holds while you increase spend, the business is scaling. If it falls, something is wrong somewhere and it is now your job to find out where.

That last clause is also the weakness. MER breaks in three specific ways.

It moves when the revenue mix moves, for reasons unrelated to ads. A large email campaign, a wholesale order, a press hit or a subscription cohort maturing all raise MER without a single thing changing in the ad account. Compare MER across two periods with different mix and you are comparing two different businesses.

It can't be compared between brands. A brand where 60% of revenue is repeat and one where it is 10% will report wildly different MER at identical acquisition efficiency, which is why published MER benchmarks are close to meaningless.

And it names no culprit. MER falling tells you something is wrong. It never tells you whether the answer is a channel, a creative, a landing page or a competitor bidding harder.

aMER fixes the first two by narrowing the numerator to new-customer revenue and the denominator to acquisition spend. It is harder to compute — you need reliable first-order flags and a defensible split of spend into acquisition and retention — and it is the number that connects directly to your unit economics, because your break-even aMER is the same 1 ÷ contribution margin from the break-even ROAS piece.

What does an incrementality test cost now?

A lot less than it did. At Google Marketing Live on 22 May 2025 Google cut the minimum spend requirement for an incrementality experiment to $5,000, from what it described only as substantial budgets. That's the single change that moves causal measurement from something enterprise teams do to something a mid-size brand can put in a quarterly plan.

Understand what you're buying, though. A platform-run holdout is the platform measuring itself with a much better instrument than an attribution window — a real control group rather than a model — but it is still the platform running the exam. An independent geo test avoids that by pausing spend in matched markets and comparing revenue, and it costs almost nothing in cash and a great deal in discipline, because the whole design collapses the moment somebody unpauses the holdout after a bad week.

Budget about a month per question. The Haus tests averaged 18.6 days of exposure plus an 8.8-day post-period, and the post-period exists because conversions arrive after the ads stop. Test one variable. Two changes in one window produce a result nobody can attribute, which is the situation you were trying to escape.

What should you steer on at your spend level?

Steer on MER, diagnose with platform ROAS and decide budget with a test. The thresholds below are a judgement about where each becomes affordable rather than a published benchmark, and the ordering is the part worth keeping.

The reason platform ROAS keeps its place at every level is that it is genuinely good at one job. Comparing eight creatives inside one ad set over one window holds everything else constant, so even a biased estimator ranks them usefully. The bias is the same for all eight. It stops being useful the moment you compare across accounts, across channels or against a target derived from your P&L.

What is wrong with the evidence in this article?

Four things, and they are worth stating because a measurement article that doesn't audit its own sources is doing the thing it complains about.

Haus sells incrementality measurement. A dataset showing that attribution windows misprice media supports the author's business, and that has to be said out loud. What makes it worth citing anyway is that the headline finding cuts against the commercially convenient story — a vendor writing marketing copy says platforms inflate their numbers, not that Meta undersells itself by 15%.

The sample skews large. Advertisers averaging $14M a year are not the median brand reading this, and a $40k-a-month account is not represented in those 640 experiments. Directional findings probably transfer. Magnitudes may not.

The DTC readout undercounts. In the same dataset, 32% of Meta's total measured impact landed outside DTC revenue, and upper-funnel campaigns showed 46% lower DTC incremental ROAS alongside a 138% omnichannel halo. If you sell only on your own site with no retail or marketplace presence, most of that 32% is not yours to claim — which cuts the other way and is exactly why an average is a poor substitute for your own test.

And the measurement ground keeps moving. Meta has narrowed its attribution windows repeatedly since the iOS 14 changes, most consequentially by retiring the seven-day view-through window, so a ROAS benchmark gathered under an older window is not measuring the same thing yours is. We are deliberately not putting dates on those changes: Meta documents them inconsistently and the ones in circulation did not survive checking. The pattern isn't Meta-specific either. Klaviyo announced in August 2024 that its default would become five-day click and five-day open, which means a passive open with no click can claim a sale. Every platform grades itself on a window it chose. Twelve ecommerce benchmarks with no source behind them works through what that does to the widely quoted share-of-revenue figures.

None of which argues for measuring less. It argues for knowing which number you are looking at, and for spending $5,000 on an answer rather than another quarter arguing about a dashboard.