Should online shops let customers keep a faulty item for part of their money back? What a Marketing Science field experiment with 853,521 shoppers found
Returns are expensive to ship, check and throw away. A huge experiment on a worldwide marketplace tested a third option between a full refund and nothing: keep the item, get part of the price back, and let an AI decide the amount.
Fang et al. tested keep-it partial refunds on 853,521 online shoppers: the first return cost more, but later spending left a net gain of $1.44 per customer.
Marketing in 3 · Issue 44 · Monday 2 November 2026
The parcel stays, part of the money comes back, and nobody queues at the post office. Photo: Rifki Kurniawan on Unsplash.
The short answer
Yes, if the refund is moderate. On a cross-border marketplace, 853,521 customers who started a return were split at random: half could be offered part of their money back to keep the item, half got the usual return. Over the next month the offer was worth $1.44 more per customer, mostly because those customers came back and spent more. Generous refunds cost money up front and invited extra claims. About 41% of the price was the sweet spot.
Evidence level
★★★One study
★★★Several studies
★★★Meta-analysis
Key takeaways
At the first return the platform paid out $0.63 more per customer and saved $0.35 on shipping and handling, so it was $0.28 down at that point.
Customers who could get the offer spent 2.2% more over the following 30 days, about $1.92 each, and the lift had not faded after 90 days.
Follow-up return requests rose by 15.5%, a sign that some customers tested how far the new option would go.
With refunds averaging half the price, the first return lost money. Cut to an average of 30%, it came out slightly ahead.
What the researchers did
Lu Fang of Zhejiang University of Finance and Economics, Chunhua Wu of the University of British Columbia's Sauder School of
Business, Zhe Yuan of Zhejiang University and Kaifu Zhang, who works at the partner platform, wanted to put a price on a
return option that big retailers are starting to try. Instead of a full refund after the item comes back, or a full refund
for cheap items that never come back, the customer keeps the product and gets part of the money back.
They ran the test on a cross-border shopping site that sells in more than 100 countries. Before the test, only items under
about $10 could be refunded without being sent back, and those got their money back in full. The new option covered dearer
items too and paid a share of the price that depended on the complaint and the evidence.
Each customer was placed in a group before the experiment began. When someone in the test group asked to return an item for
a reason that could be checked, such as damage or a wrong colour, three AI agents read the message, looked at the photos or
video, and either made an offer or sent the case down the normal route. Customers could take the offer or ship the item back.
They could not haggle. The platform also made the offers less generous partway through, which let the team see what
the refund level itself does.
Customers
853,521 people who started a return between August 2024 and January 2025
Groups
425,578 could receive a keep-it offer; 427,943 had only the usual return
Products
The platform's own stock: lifestyle, fashion and small electronics, about $27 per item on average
Offers
Fixed shares of the price set by complaint type and severity, averaging 50% in the first phase and 30% in the second
Not included
Change-of-mind returns such as "no longer needed", which an AI cannot check against evidence
Tracking
Each customer's spending and returns for up to 90 days after the first request
Who took the offer
Between 60% and 70% of customers in the test group received an offer at their first return. The rest either sent unclear
evidence or had their claim turned down, and went through the normal process without knowing an AI had looked at it.
How people responded depended on the complaint. A crushed box or a scratch is easy to see and easy to accept, and around four
in ten of those offers were taken. Complaints about quality, style or a colour that looked different online were harder to
settle with money alone. Acceptance there was under three in ten. When the platform cut the average refund from about half
the price to under a third, acceptance dropped by roughly ten percentage points.
The new option mostly changed what happened to people who would otherwise have given up. Against the usual process, returns
ending with no refund at all dropped by 14.8 percentage points, and returns ending with a parcel sent back dropped by 6.8. That is why payouts went up: many people who would have walked away now took some money instead.
Share of keep-it offers that customers accepted, by return reason (%)
Share of keep-it offers that customers accepted, by return reason (%)
Group
Value (%)
Damaged packaging
43.4
Scratched
41.2
Arrived damaged
39.8
Product flaw
36.6
Poor quality
32.0
Colour not as described
29.8
Does not work properly
29.2
Appearance not as described
23.4
Style not as described
22.3
Test group, offers made at the first return, both phases. Amber: visible defects; grey: complaints about how the product seems.
The first return costs money, the next month pays it back
How generous the offer was decided whether the first return made or lost money. In the generous phase the platform was $0.90
per customer worse off at that point. In the leaner phase it came out $0.10 ahead, because the savings on shipping and checking
returned goods outweighed the small rise in payouts.
The bigger effect came later. People in the test group were 1.2% more likely to buy again within a month, placed 1.7%
more orders and bought from 1.6% more product categories. Within the group that received an offer, raising the refund by ten
percentage points was linked to about $1.20 more spending in the following month.
Put together, the platform gained in net terms, though the margin matters. Assuming a 35% gross margin, the profit gain was
about $0.19 per customer over 30 days, or $0.62 in the leaner phase. Over 90 days the net gain reached $4.41 per customer, about
$1.23 in profit. Fewer parcels also went back across borders: ship-backs fell by about 12%, which the authors equate to the
carbon soaked up by roughly 2,000 mature trees across the customers in the experiment.
Some customers started testing the system
The rise in later return requests was not all goodwill. The AI's rules stayed fixed throughout, so any change in how often it
approved claims reflects the claims, not the judge. Approvals for follow-up requests slipped compared with first requests, and
the drop was biggest exactly where bending the truth is easiest and pays most: vague quality complaints, expensive items and
repeat claims made within a week.
A second clue came when the AI said no. Normally a genuine complaint is worth the effort of posting the item back. For high-priced
items, the share of rejected follow-up claims that ended in a parcel being sent back fell by 6.34 points, which looks
like people trying their luck and walking away when it failed. Returned orders rose by 4.3%, nearly twice the rise in spending.
More generous first offers made this probing more likely.
Fall in AI approval rate for follow-up claims compared with first claims (percentage points)
Fall in AI approval rate for follow-up claims compared with first claims (percentage points)
Group
Value ()
All follow-up claims
1.69
Visible defects
1.49
Quality or looks complaints
2.18
Lower-priced items
0.85
Higher-priced items
3.02
Repeat claim within a week
3.46
Repeat claim after a week
0.65
Test group, claims within 30 days of the first return. The AI's rules stayed fixed, so the drop reflects the claims being made.
How generous should the refund be?
The two phases let the team model what each refund level is worth. Judged only on the first return, the best average refund
would be just 6% of the price, and the option would stop beating a normal return above 36%. Once later spending is counted,
the best level rises to about 41%, and the option keeps beating a normal return all the way up to 90%. Both real phases, at
50% and 30%, sat inside the range where the scheme paid.
Scale was the other condition. A person needs about five minutes to weigh up a claim with photos. The three AI agents took
about five seconds, and the authors put the cost of doing the same work by hand for this sample alone at over half a million
dollars. So these results describe a refund rule and the AI that made it workable, tested together.
Average refund (% of price) that earns the most, and the highest refund that still beats a normal return
Average refund (% of price) that earns the most, and the highest refund that still beats a normal return
Group
Value (%)
Best refund, first return only
6
Best refund, with next month's spending
41
Break-even, first return only
36
Break-even, with next month's spending
90
The authors' calibration from the two phases of the experiment (average refunds of 50% and 30%).
Marketing in 3 steps
Ready. Set. Market.
1
ReadyDecide and prepare
List where returns cost you most. Pull last quarter's returns and, for each product line, write down the average item price next to what it costs you to get one back, check it and restock it. Mark every line where getting it back costs more than a third of what you can resell it for.
Decide what proof you will accept. For each return reason on those lines, write one sentence on the evidence that settles it, such as a photo of the damage or a short video of the fault. Leave change-of-mind returns on the normal route.
2
SetBuild it
Build a simple refund table. Set a fixed share of the price for each reason and how bad the problem is. Start in the middle of your range, and pay less for complaints that are hard to verify, like style or quality, than for visible damage.
Make one clear offer. Show the keep-it amount right beside the normal return option, in plain words, with one button for each. Do not let customers bargain, and keep the table the same for everyone in the trial.
3
MarketLaunch and measure
Trial it on half your returners. For eight weeks, give the option to a random half of customers who start a return and keep the other half on your usual process. Compare refunds, return costs and what both groups spend over the next 30 and 90 days.
Flag repeat claims early. Each week, list customers who file a second claim within seven days, claims on your priciest items, and claims dropped right after a refusal. If those lists grow, lower the refund for those cases before you widen the trial.
Found this useful?
Marketing in 3 is free and ad-free, with new issues every Monday and Thursday. A coffee keeps it that way.
A dip before the gain. The return itself usually costs a little more, so judge the scheme on what customers do in the weeks after.
Fewer abandoned returns. People who would have given up and stayed annoyed now get something, and that is where the goodwill comes from.
Value for shoppers too. If people keep items worth about half their price, the authors estimate they hold on to about $1.49 each, on top of the refund.
More claims to screen. A smoother refund draws more requests, so budget for checking them and expect some to be opportunistic.
Where it does not work
One platform, cheap goods. Items averaged $27 and returned stock had almost no resale value. Where returns can be resold, such as furniture or large electronics, the sums change.
Rule and AI tested together. There is no way to split the result between the refund itself and the speed of the AI decisions.
A quiet test. The option was not advertised and reached under 1% of the platform's users, so word of mouth and rival moves were left out.
Profit is an estimate. The profit figures assume a 35% margin; the measured outcome is spending minus refunds and handling costs.
Questions marketers ask
Do returnless partial refunds make money for online retailers?
In this experiment, yes, once later spending is counted. The first return cost slightly more, but customers offered the option bought more over the next month, leaving a net gain of $1.44 per customer after 30 days and $4.41 after 90 (Fang et al., 2026).
How big should a partial refund be?
The authors' model puts the most profitable average at about 41% of the item's price. Judging only by the first return would push it far too low, to around 6%.
Will customers abuse a keep-it refund?
Some will try. Follow-up claims rose, approvals fell most for vague complaints, pricey items and quick repeat claims, and more customers dropped a claim once it was refused. The overall result stayed positive, but the abuse grows with generosity.
Why would a partial refund make people spend more later?
It turns a hassle into a quick fix. Customers who could resolve a problem without posting anything back bought again more often and across more categories, which is what research on service recovery would predict.
Can AI decide refund claims on its own?
Here it did, for claims backed by photos, video or text. Three AI agents recognised the request, checked the evidence and picked an amount from a fixed table in about five seconds. Unclear claims went to the normal return process.
Using shopper-level data from an Asian online platform, this study finds that opening self-pickup stations raised spending and shopping frequency without eating into home delivery. The gain came from keeping existing customers, not from winning new ones. It shrank when a station sat close to big physical stores, where shoppers could simply buy elsewhere. Location is part of the offer.
Text mining of nearly 372,000 complaint and reply pairs for hotels and restaurants on Tripadvisor and Yelp links better replies to more positive reviews afterwards. Replies that pick up the specific service points raised did best, especially when they also signalled fair process or compensation. Replying to selected complaints worked better than answering every negative review.
Source
Fang, L., Wu, C., Yuan, Z., & Zhang, K. (2026). Returnless Partial Refunds: A Large-Scale Field Experiment in E-Commerce. Marketing Science. https://doi.org/10.1287/mksc.2025.0618
This brief is written for practitioners from the published article. Open access (CC BY 4.0): free to read at the publisher. Numbers are as reported by the authors; read the article for full methods.
Facing all-White staff, 50% of non-White customers chose a digital kiosk, against 36% with non-White staff. What it means for service design and AI options.
Bruno Schivinski is Associate Professor of Marketing at Gdańsk University of Technology. He studies how consumers engage with brands in digital and social media.
