Your highest-ROAS campaign may be your highest-return campaign
Bidding models learn at checkout. Returns arrive weeks later, and the model never hears about them.
By Nathan Barling Product Marketing Principal 5 min read
Figure 1
Same campaign, two ROAS numbers
Illustrative: a 4x reported ROAS with a 35% return rate
Bar chart: 4.0x reported ROAS falls to 2.6x after a 35% return rate
Retailers expected 15.8% of 2025 sales to come back, and online orders came back more often, at 19.3%, according to the NRF and Happy Returns. An ad platform counts each of those sales at checkout. It never hears about the box that comes back three weeks later.
Picture two campaigns, both reporting a 4x return on ad spend. One sells basics that people keep. The other sells items people order in three sizes and send two back.
In the ad platform, the two look the same. Each sale was counted at checkout. Neither platform heard about the boxes that came back three weeks later.
That's the whole problem. A bidding model optimizes the value you send it. If you send gross checkout value, it gets better at finding people who buy, not people who keep what they buy. So your highest-ROAS campaign can be your highest-return campaign, and the model has no reason to prefer the other one.
What the bidding model sees
A purchase event fires at checkout with an order value. Refunds and returns live somewhere else: an order management system, a returns portal, finance. Unless someone connects them, the model keeps treating the original order as a win.
The platforms do offer ways to correct this. In Google Ads, a retraction removes a conversion and a restatement changes its value, and the API documentation describes matching them by order ID. Meta's Conversions API guide describes a net_revenue parameter that sends profit alongside value on a Purchase event. That guide doesn't say how refunds are handled, so confirm before you rely on it.
Timing is the other half. A returns platform found that 80% of returns arrived within 14 days in an earlier analysis, and within 21 days in a newer one. That's 2021 data from one vendor, so treat it as direction, not a benchmark. And a marketplace analytics vendor says most teams review returns after finance closes the month. That's a practitioner claim, not data, but it matches how the pipes usually look.
What it does to your best campaign
The arithmetic is plain. Take a 4x reported ROAS and a 35% return rate. The revenue the customer kept supports a 2.6x return, before margin. If your best campaigns skew toward categories that return often, the distortion is largest exactly where you're most pleased with the results.
How often is that? The National Retail Federation and Happy Returns projected that 15.8% of 2025 sales would be returned, down from 16.9% the year before, with online higher at 19.3%. Total returns were projected at $849.9 billion. Both organizations have a stake in returns services, and these are survey projections from October 2025, the latest edition I could find. Still, if you sell online, nearly one order in five can come back.
What closing the loop looks like
Fashion group Bestseller published a case study with Google in April 2026. The team predicted the likelihood of a return for each order at checkout, subtracted it from the order value, and sent the adjusted number to Google Ads for bidding. In Google's example, a $100 order with a predicted $60 return goes in as $40.
Google reports up to 24.5% lower cost per click in tests in the Netherlands and Sweden, a 50% higher ROAS in the Netherlands (on the adjusted value basis), and 93% accuracy in return predictions. The page doesn't define that accuracy figure.
Read all of it as a vendor-reported case. Google published it and sells the products involved, the figures say "up to," and the page says results vary by advertiser. It's also a prediction, not a reconciliation with finance.
The detail worth keeping is a failure. The team's first plan was to exclude frequent returners, and it didn't work because so many shoppers were new or anonymous. Predicting at the order level worked where targeting at the customer level didn't.
Three ways to send a better value
Predict
Estimate return likelihood by product, category, or customer at checkout and subtract it from the value you send. Fast, and only as good as the model.
Correct
When a return is logged, upload a retraction or restatement matched to the order ID. Accurate, but it lands late.
Reconcile
Send net revenue or profit values that finance has signed off on. The most meaningful signal, and the hardest to agree on.
A reasonable setup uses the first two together: a prediction to steer bidding now, and a correction when the real number arrives. Waiting for the truth makes the signal accurate and late, since the platform has already spent the money. Predicting makes it fast and sometimes wrong. So pick the shortest delay your data supports, predict inside it, and compare predictions with actual returns every month.
Keep your own reporting on the net figure either way. A platform that sees a smaller number than you do isn't the problem. A team that judges campaigns on the gross number is. Someone also has to own the definition of net revenue, and writing it down once in a metrics layer keeps every report and every upload on the same formula.
One caution. A high return rate doesn't make a campaign bad. Some categories return often and still earn healthy margins. The aim is to optimize to what the business keeps, not to the lowest return rate.
Where Alli fits
Connecting returns data to ad data is mostly a data foundation job. Alli's Unified Data Layer centralizes platform and regional data into one governed foundation with traceable data lineage. Categorizations apply consistent rules to product and campaign labels, which makes return rates comparable across markets and platforms. Alli Audience Gateway delivers site events to Meta, Google, TikTok, and Pinterest from one implementation, using server-side tracking.
None of that decides which values should go to the platforms. That definition, and who approves it, stays with your team. Whether a platform accepts later adjustments also depends on that platform.
When this isn't the problem
If few of your orders come back, or you sell something where gross and net are close, the effort may not pay. If tracking is dropping conversions before they reach the platform, repair that first. Signal loss means a corrected value on a missing conversion helps no one. And a net value tells the platform what a sale was worth, not whether the ad caused it. That's a question for incrementality testing.
Pick one platform and one high-return category. Compare the value you send with the value finance books three weeks later. The difference is your first number.
Terms in this article
- Conversions API
- Server-to-server conversion event delivery to ad platforms, independent of browser signals.
- Server-side tracking
- Sending event data from a company's server to ad platforms, bypassing browser limits.
- Data lineage
- A record tracing each number back to its source and every transformation step.
- Metrics layer
- Defining each KPI once, with its formula and filters, so every tool calculates it the same way.
- Signal loss
- The share of customer actions that measurement tools never see.
- Incrementality
- The extra result marketing caused beyond what would have happened anyway.
Frequently asked questions
Why can ROAS be misleading?
ROAS divides the revenue an ad platform recorded by ad spend. If the platform records the order at checkout and never hears about the return, ROAS counts revenue the customer didn't keep. Campaigns that sell frequently returned products look better than they are.
What is return-adjusted ROAS?
It's ROAS calculated on revenue after returns, ideally after margin. A 4x reported ROAS with a 35% return rate becomes 2.6x on kept revenue. It tells you what a campaign earned, not what it was credited with.
How do you send return data back to Google Ads?
Google Ads supports conversion adjustments. A retraction removes a conversion, and a restatement changes its value. Orders are matched by order ID. You can also send a return-adjusted value at checkout. Check Google's current documentation for time limits that apply to your account.
How long do returns take to show up?
It depends on the retailer. One returns platform found that 80% of returns arrived within 14 days in an earlier study and within 21 days in a newer one. That analysis is from 2021, so measure the lag in your own order data.
Does Meta support profit optimization?
Meta's Conversions API guide describes a net_revenue parameter on the Purchase event that sends profit alongside value. The guide doesn't cover refunds or returns, so confirm with your Meta team how later adjustments are handled before you rely on it.
Is predicting returns accurate enough to bid on?
Bestseller reports 93% accuracy in return predictions in a Google case study, though the page doesn't define the metric. Treat that as vendor-reported. Compare predictions with actual returns every month and retrain the model when they drift.
Sources
- NRF and Happy Returns, Consumers Expected to Return Nearly $850 Billion in Merchandise in 2025, October 15, 2025. Happy Returns sells returns services.
- Google (Think with Google UK), Stop paying for returns: Inside Bestseller's AI system that knows what's coming back, April 8, 2026. Google sells the products described.
- Google Ads Help, About conversion adjustments. Google sells Google Ads.
- Google Ads API, Upload conversion adjustment, updated September 30, 2026.
- Meta for Developers, Value optimization for profit: integration guide. Undated. Meta sells ads.
- Loop Returns, How long do customers take to return?, May 2021. Older data. Loop sells returns software.
- FiveX, Marketplace returns forecasting in 2026, July 2026. FiveX sells marketplace analytics.
- JudeLuxe, Beyond ROAS, December 26, 2025. Agency; hypothetical example only.
Last updated November 6, 2026. Statistics are quoted from the sources above and were checked on that date. Survey results are self-reported unless noted.