Why ROAS falls when you scale spend (and what to watch)

The extra spend has to buy conversions that cost more than the first spend did, and both platforms say so. Meta's bid strategy documentation says that with its lowest-cost strategy, costs "can rise as you exhaust least expensive opportunities or as you increase budget", and Google's says that to get more conversion volume from a target ROAS you reduce the target so the bidder can "enter more auctions". The number to watch is marginal ROAS, the return on the next euro, and a step-up test in one region reads it.

Why does the average fall when you raise the budget?

A platform goes to the cheapest conversions first and then reaches further, so the next euro can return less than the first ones did. ROAS columns are averages: Google Ads calculates conversion value per cost "by dividing your total conversion value by the total cost of all ad interactions", one number over all spend, not the return on the last euro.

For illustration: spending 1,000 attributes 4,000 of revenue (a ROAS of 4.0), and raising spend to 1,500 attributes 5,000, so the average falls to 5,000 / 1,500 = 3.3 while the extra 500 of spend returned (5,000 - 4,000) / 500 = 2.0. For illustration: with a 40% contribution margin, break-even ROAS is 1 / 0.4 = 2.5, so the extra 500 lost money while the average still looked comfortable.

Field experiments show how the weak spend can hide inside an average. In eBay's experiments on non-brand keywords (Blake, Nosko and Tadelis, Econometrica, 2015), new and infrequent users were positively influenced by ads, but more frequent users, whose purchasing behavior was not influenced by ads, accounted for most of the advertising expenses, and average returns were negative. That is a snapshot of one budget on one platform, not a measurement of a budget step. No public measurement of the return on a budget step for Shopify brands turned up in the sources read for this page, so the step-up test below is how you get yours.

Scaling isn't a mistake. A step can be worth taking while the average falls, as long as the return on that step clears break-even.

What is a saturation curve, and what do marketing mix models assume?

Two open-source marketing mix modeling tools, Meta's Robyn and Google's Meridian, both build diminishing returns into the model as an assumption. Robyn's documentation says the theory of diminishing returns, also called saturation, "holds that each additional unit of advertising investment increases the response at a declining rate", and implements it as a variable transformation, a Hill function whose parameters set the curve's shape and inflexion point. Meridian's documentation defines ROI as "a historical, channel-wide average" and mROI as "the return on the next dollar spent", and says that if the mROI is "much lower compared to the ROI", the channel is beginning to saturate at historical spend level.

Both also warn against trusting a curve past your data. Meridian says the fitted response curve can be extrapolated outside the observed range but "results based on extrapolation should be interpreted with an appropriate level of caution", and that the risk "increases the further out you go". A curve fitted to last year's spend says little about a step well above it. A step tells you directly.

How do you read marginal ROAS with a step-up test?

1. Split regions at random into two groups of similar order volume. Google's own geo experiments randomly assign non-overlapping regions to a control or treatment condition.
2. Raise spend once, in one group, by a fixed step, and leave the other group alone. Change nothing else, and don't keep adjusting: Google says frequent manual changes to budgets or targets reset the standard 7-14 day learning period, and that after a target change you should give the bidder 1-2 conversion cycles. Google also says its bidding changes from 17 August 2026 for budget-limited Target CPA and Target ROAS campaigns "may cause temporary performance and traffic fluctuations", so pick your before weeks with care.
3. Pull Shopify net sales by region for equal numbers of weeks before and during the test. For each group divide revenue during by revenue before. The gap between the two ratios, times the stepped group's revenue before, is the revenue the step added. Divide it by the extra spend: that is marginal ROAS.
4. Compare it with break-even ROAS, one divided by your contribution margin, using the break-even ROAS calculator. Above it, the next step is worth testing. Below it, hold or step back, whatever the average says.
5. Mind the noise and the range. In Lewis and Rao's 25 field experiments the median confidence interval on return on investment was over 100 percentage points wide, so a small step in a small store can come back inconclusive. Use a bigger step or more weeks, and move one step at a time.

Pass: marginal ROAS above break-even. Fail: below it. Inconclusive: the gap sits inside normal week-to-week variation, which means not measured, not flat.

Between tests, watch blended MER, total revenue divided by total ad spend, next to platform ROAS: it counts each order once, so no platform can claim a sale another one already took.

Where does a read fit?

Later, once a step or two has told you where a channel stands, a causal attribution read like Causality Engine's gives a next step per channel, including how to test it, from a GA4 export.

Sources, 30 September 2026: Bid Strategy (Meta for Developers, 2026); About Target ROAS bidding (Google Ads Help, 2026); Understand your conversion tracking data (Google Ads Help, 2026); Robyn and Key Features (Robyn documentation, Meta, 2026); Meridian (Google for Developers, 2026); Media saturation and lagging (Meridian documentation, Google, 2026); Incremental Outcome, ROI, mROI & Response Curves (Meridian documentation, Google, 2026); Measuring Ad Effectiveness Using Geo Experiments (Google Research, vendor-published); The Unfavorable Economics of Measuring the Returns to Advertising (Lewis and Rao, Quarterly Journal of Economics, 2015, via EconPapers); Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment (Blake, Nosko and Tadelis, Econometrica, 2015, via EconPapers).

Related answers

• What a budget ladder does to incremental ROAS
• Meta Ads ROAS dropped: what changed and what to check
• Google Ads ROAS dropped: a seven-point diagnosis
• Blended MER vs platform ROAS: which should drive spend?
• Marketing mix modeling, explained