AI Size Recommendations: Measure Them Against Your Own Returns
Fit recommendation tools promise a lot and report their own success. How to test one against your returns data by reason code, size run, and category before renewing anything.
The vendor dashboard says the size recommendation tool is working: engagement up, confidence high, a conversion lift on sessions that used it. None of those numbers answer the only question the tool was bought to answer, which is whether fewer customers are receiving the wrong size and sending it back.
That answer lives in your returns data, not the vendor's, and it takes an afternoon to pull. Stores skip the exercise because the dashboard is easier to read in the renewal meeting. Then they renew a tool that changed buying behaviour without changing fit outcomes, and the returns rate, the cost line the tool was meant to move, sits exactly where it was.
Slice returns by reason, not by total
Start with returns coded too small and too large as a share of orders, for customers who used the recommendation against a comparable group who did not, across the same period and the same categories. The total returns rate will mislead you: it mixes damage, changed mind, late delivery, and bracketing, none of which a fit tool can influence. The size-reason slice is the tool's actual report card.
Watch the direction as well as the volume. A tool that systematically recommends a size up will cut too small returns and grow too large ones, and the vendor dashboard will call the first half a win. Both halves are yours. Also split by category before concluding anything: a tool trained mostly on one range, say core denim, can be accurate there and worse than nothing on footwear, and the blended figure will hide the footwear problem behind the denim success.
Run the test the vendor would not choose
Compare recommendation users with non-users carefully, because the groups differ before the tool touches them. Shoppers who use a fit tool are often the uncertain ones, between sizes, new to the brand. Their baseline return rate is higher. A fair read compares like with like as far as you can: same categories, similar order values, first time against repeat buyers separated. It will not be a randomised test. Say so in the write up, and let the size of the gap, not false precision, carry the decision.
And check what the tool recommended against what it should have. Take fifty completed orders where the customer kept the item, look at the size bought and whether the tool agreed. Then fifty size returns, and ask whether the tool pointed at the size that went back. That manual sample catches a tool that is confidently, consistently one size off in a category, which aggregate rates can take months to confess.
Decide what the tool has to beat
Before the renewal conversation, write down the bar: the size-reason return reduction that justifies the tool's cost per order against the same money spent on better size charts, model measurements, or review mining for fit notes. Tools survive renewals on narrative when no bar was set in advance. A tool that clears a stated bar is a good tool even if its dashboard is ugly, and one that cannot is a subscription to a chart. Set the bar while you are still allowed to be unimpressed.
Related reading
If this topic is useful, these earlier guides go deeper on connected questions: AI-Translated Product Pages: When Sizes, Care, and Safety Wording Go Wrong, When AI Search Answers and Your Product Feed Disagree, and AI Review Summaries Smooth Away the Patterns You Actually Need.
Frequently asked questions
What metric actually judges a size recommendation tool?
Size-reason returns, coded too small and too large, as a share of orders, comparing tool users with comparable non-users in the same categories and period. Total returns mix in reasons the tool cannot affect.
Why not trust the vendor dashboard?
It reports engagement and conversion on sessions that used the tool, which are self-selected shoppers. It rarely measures kept orders or size-reason returns, the outcomes the tool was bought to change.
What if the tool recommends well in one category and badly in another?
Common, especially where training data is deep in one range and thin in another. Split results by category and be ready to switch the tool off where it underperforms a good size chart.
How large a sample do you need for a manual check?
Fifty kept orders and fifty size returns, comparing the bought size with what the tool recommended, is enough to spot a consistent one-size bias in a category.