Measuring Returns After Virtual Try-On
Measure fashion returns after try-on with reason codes, order-date cohorts, mix controls, and net margin, not a store-wide rate that hides mix shifts on Shopify.
To measure whether try-on reduces fashion returns, stop reading the store-wide rate. Define reason codes that separate appearance from size, cohort orders by purchase date, hold category mix still, and report net margin after refunds and inbound freight. A preview can change what comes back. It cannot prove the garment would have fit in a fitting room.

Reason codes on paper, garments on the table. Editorial image in Classic Antla disposable-camera style.
A store-wide return rate is a blended number. It mixes trousers that always come back, scarves that almost never do, a sale week, a new size run, and whatever you changed on the product page. If you turn on a preview widget and then stare at that blend, you will invent a story. The work below is how to stop inventing.
Shopify already stores the raw material. Returns and exchanges attach a reason to each line. Your job is to treat those reasons as a dataset, not a customer-service afterthought.
Define return reason codes
Most fashion stores run on three tired labels: too small, too big, and not as described. Those labels collapse two different failures. Size failures are measurement and grading. Appearance failures are color, drape, length, and “looked different on me.” A preview is an appearance tool. If you dump both failures into “fit,” you cannot tell whether the widget did anything.
Rebuild the reason list so a warehouse person can pick one without guessing. Keep it short. A twelve-item taxonomy that nobody uses is worse than four items used every day.
| Code | Means | Preview can affect? |
|---|---|---|
| Size: small / large | The shopper says the garment does not match the size they expected | Weakly. Preview is not a tape measure |
| Appearance: color | The color on body or in daylight is not what the listing showed | Yes |
| Appearance: silhouette | Length, drape, or cut looks different on them than on the model | Yes |
| Quality / defect | Stitching, pilling, odor, damage | No |
| Changed mind | No product issue | No, unless the listing oversold the look |
| Other / blank | Unusable | Fix the form |
Train the warehouse, not just the theme. If the default is blank, you will get blanks. If “Other” is first in the list, you will get Other.
Pair each code with a photo when the shopper uploads one. You are not building a court case. You are checking whether “looked different” clusters on one fabric or one photographer.
Two outside numbers belong in the briefing, not in your dashboard as targets. The NRF 2024 returns research is about why people return across retail, not your SKU. Antla’s fashion return rate benchmarks and 2026 returns statistics are category context. Copying either into a board slide as “our number” is how teams lie to themselves.
Cohort by order date
Returns lag. A dress ordered on Monday can come back three weeks later. If you chart returns by the day the warehouse received them, a quiet shipping week looks like a miracle.
Cohort by order date. For every order placed in week W, count how many of those order lines were returned by a fixed horizon: 14 days, 30 days, and 45 days after delivery, or after the return window you actually offer. Then stop. Do not keep extending the horizon until the chart looks good.
Exclude orders you cannot treat as comparable:
- Wholesale and draft orders
- Exchanges that are not net new demand
- Gift returns where the buyer never saw the page
- Channels with no preview (a wholesale portal, a marketplace)
If try-on lives only on the online store, do not mix point-of-sale orders into the cohort.
Write the window on the worksheet before you launch. “We will read 30-day return rate for orders placed in the first four weeks after launch, compared with the four weeks before, same product set.” That sentence is the experiment. Everything else is decoration.
The mechanism page, how try-on can cut returns before checkout, explains why appearance confidence might change a later return. This article is only the measurement. Do not quote the mechanism as proof.
Control category mix
A drop of heavy coats next to a drop of tanks will move the store-wide return rate even if nobody changed a pixel. Category mix is the usual confounder.
Hold the mix still in the readout, not in the warehouse. You still sell what you sell. You just do not average coats and tanks into one percentage and call it a try-on result.
Practical controls, cheapest first:
- Report return rate inside each product type: dresses, denim, knits, outerwear.
- Weight those rates by the pre-period mix if you need one headline number.
- Keep a matched SKU set: items that existed before and after launch, with inventory in the same sizes.
- If you launched try-on on a subset of SKUs, those SKUs are the test. The rest are a natural control, as long as merchandising did not change only on the test set.
If you cannot hold mix, say so. A before-and-after on a store that swapped 40 percent of the catalog is not a try-on study. It is a merchandising study with a widget in the screenshot.
Bracketing makes mix worse. Shoppers who buy two sizes and send one back inflate units and returns together. The cost of bracketing is a separate ledger line. Do not hide it inside “try-on reduced returns” if the real change was fewer two-size orders.
Report net margin
A cheaper return that still kills margin is not a win. Refunds, inbound labels, processing time, and unsellable units are the costs that matter. Shopify’s return workflows tell you units. Finance tells you money.
Build a one-page margin view for the same cohorts:
| Line | How to get it |
|---|---|
| Gross merchandise of cohort | Order reports, product type filter |
| Refunds in the return window | Return reports joined to those orders |
| Inbound freight and labels | Carrier invoices, allocated by returned units |
| Processing labor | Warehouse hours on returns, even a weekly average |
| Unsellable / restock haircut | Units that cannot go back at full price |
| Net after those costs | The number you actually discuss |
Then ask whether appearance-coded returns moved. If size-coded returns moved and appearance-coded returns did not, the preview is not the story. If appearance-coded returns moved and net margin did not, you may have shifted refunds into discounts or into longer handling.
Across Antla merchants, return rates have fallen by up to 30% versus the pre-try-on period in the same stores. That compares a store with try-on live against its own earlier period, so mix, weather, and policy can still leak in. It is an association in merchant reporting, not proof that the preview caused the drop. Antla merchant reporting, cohort dates not published.
Use that figure as a ceiling for planning, not as a forecast you put in a budget. Your cohort worksheet is the source of truth for your store.
If you want a preview on Shopify product pages while you run this measurement, Antla is one option. Keep measurements and size notes. A try-on shows how a piece may look. It does not confirm physical fit.
Where this readout fails
- Policy changes. Extending the return window in the same month as launch will swamp the chart.
- Reason-code drift. If the warehouse starts using “not as described” for everything, appearance codes go to zero.
- Small volume. A type with 40 orders a month cannot support a 30-day claim. Group types or wait.
- Cross-channel. A shopper who previewed on the site and returned in store still belongs in the online cohort if that is where they bought.
- Selection. Shoppers who use try-on are not a random sample. Do not treat try-on-user return rate as store-wide impact. That is a different design, closer to a holdout.
For the confidence problem that often sits under appearance returns, see fit confidence on fashion PDPs.
Next action this week
Export four weeks of returns with reason, SKU, and original order date. Map reasons into the table above. Pick one product type with enough volume. Write the cohort sentence for that type, including the horizon. Do not change the widget until that sentence exists.
About the author: Aaron is the founder of Antla and spends an unreasonable share of his week reading fashion product pages on a phone.