On this page
Every size recommendation vendor puts a big number on their homepage — 90%, 95%, sometimes 98% "accuracy." Almost none of them define what that number actually measures, and the gap between a top-1 accuracy claim and a within-one-size accuracy claim can be the difference between a return-rate win and a return-rate disaster. If you are evaluating a size recommender, or already run one, you need to understand exactly what you are buying before you trust the badge on the checkout page.
What "Accuracy" Actually Means in Fit Prediction
Size recommendation accuracy is not one number. It is a family of metrics, and vendors pick whichever one flatters them most.
Top-1 Accuracy
Top-1 accuracy asks a strict question: did the single size the system recommended match the size the customer actually kept without returning? If your recommender says "Medium" and the customer keeps a Medium, that's a hit. If they keep a Small or a Large, it's a miss — no partial credit. This is the honest metric, and it is almost always lower than what marketing pages advertise.
Within-One-Size Accuracy
Within-one-size accuracy is more forgiving: it counts a recommendation as correct if the kept size is the recommended size OR one size up/down. A system recommending Medium counts as "accurate" if the customer keeps Small, Medium, or Large. This metric inflates numbers substantially — a recommender with 68% top-1 accuracy might report 91% within-one-size accuracy, and both statements can be technically true.
The problem: within-one-size accuracy is a poor proxy for what actually drives returns. A customer who gets a Small when they needed a Large is still going to return the item, even though the model "got close" in categorical terms. For apparel with tight sizing curves — denim, structured outerwear, fitted knitwear — a one-size miss is a full return, not a near-hit.
Realistic Benchmarks You Should Expect
Published academic and industry work on fit prediction (using purchase and return history from large retailers) generally lands top-1 accuracy in the 60-75% range for well-tuned models on categories with rich sizing data, and within-one-size accuracy in the high 80s to low 90s. Machine learning researchers building on retail return datasets have repeatedly found that fit prediction is a genuinely hard multi-class problem — body shape, garment cut, brand-specific sizing drift, and fabric behavior all interact — which is why claims above 90% top-1 accuracy, without qualification, should be treated skeptically. See general discussion of accuracy metric definitions and their limitations in classification problems from sources like McKinsey's retail analytics research.
| Metric | Typical range (mature systems) | What it really tells you |
|---|---|---|
| Top-1 accuracy | 60% - 75% | Exact size match rate — the number that predicts returns |
| Within-one-size accuracy | 85% - 93% | How often the model is "in the neighborhood" |
| Category-specific accuracy (denim, fitted) | Often 10-15 points lower than knitwear | Tight-tolerance categories punish near-misses |
| New-customer (cold start) accuracy | 15-25 points lower than repeat-customer | Lack of purchase history hurts most models |
Cold-start accuracy — predictions for a first-time shopper with no purchase history — is the number vendors rarely disclose, and it is often the weakest part of the system. If a large share of your traffic is new visitors, ask for this figure specifically.
How to Measure Accuracy Yourself, Not Trust the Vendor's Number
Never accept a vendor accuracy claim without asking three questions:
- Which metric is this — top-1 or within-one-size? If they can't answer clearly, assume within-one-size, because that's the flattering one.
- What's the test set? A number computed on the vendor's best-performing client category (say, basic t-shirts) tells you nothing about your fitted-dress or denim assortment.
- What's the cold-start number? If they only report accuracy for repeat customers with size history, you're seeing a best-case scenario.
Run Your Own Holdout Test
The most reliable approach is to run a live A/B or holdout test on your own store for 60-90 days: expose a portion of traffic to recommendations, hold back a control group, and compare actual return rates and exchange rates between the two — not the recommender's self-reported confidence score. This is the only accuracy number that matters for your P&L, because it accounts for your specific product catalog, sizing charts, and customer base.
Track the Metric That Matters: Size-Related Return Rate
Recommendation "accuracy" is really a proxy metric. The metric you actually care about is size-related return rate — the share of returns where the stated reason is "too big," "too small," or "doesn't fit." Watch this number before and after deploying a recommender, segmented by category. A tool with modest top-1 accuracy that still cuts size-related returns by 20-30% is doing its job; a tool with an inflated accuracy badge that doesn't move your return rate is not.
If you're comparing vendors, this is a core part of how to choose a size recommendation solution — accuracy claims should never be the first filter, real return-rate impact should be.
Why Category and Data Depth Change the Number
A recommender's accuracy is not a fixed property of the algorithm — it moves with category and data availability.
Category Complexity
Simple, boxy garments (t-shirts, basic sweatshirts) tolerate a wider margin of error before a return happens, so accuracy numbers on those categories look artificially strong. Fitted, structured, or size-critical categories (bras, tailored blazers, jeans) have tighter tolerance, and accuracy on those categories is the real stress test of a system.
Data Depth Per SKU
Recommenders trained on brands with thin size-chart data, few reviews, and low purchase volume per SKU will underperform regardless of the underlying model quality. If you're a smaller merchant evaluating vendors, ask how they handle new or low-volume SKUs specifically, since this is where most accuracy gaps show up in production, not in the vendor's demo.
This connects directly to overall ROI: a mismatch between advertised accuracy and your actual category mix is one of the most common reasons the ROI of a size recommendation tool underdelivers relative to projections. Before signing a contract, model the expected return-rate impact using your own category breakdown, not a vendor's blended average.
Setting Internal Accuracy Targets
Rather than chasing a headline percentage, set two internal targets:
- Top-1 accuracy floor: don't accept a system below roughly 65% top-1 on your core categories after the first 90 days of tuning.
- Return-rate delta: require a measurable, statistically significant reduction in size-related returns within one quarter, tracked against your pre-deployment baseline.
If a vendor can't commit to reporting both numbers post-launch, that's a signal worth taking seriously during procurement.
For a broader comparison of vendors and how they report these metrics, see our roundup of size recommendation tools, which breaks down accuracy methodology alongside pricing and integration effort.
Frequently Asked Questions
What is a good top-1 accuracy for a size recommender?
For mature systems on data-rich categories, top-1 accuracy in the 65-75% range is considered strong. Anything advertised above 90% top-1 accuracy without a clear methodology and test-set description should be treated with skepticism, since that range is rarely achieved outside of narrow, low-variance product categories.
Why do vendors report within-one-size accuracy instead of top-1?
Within-one-size accuracy produces a much higher, more marketable number because it counts near-misses as correct. It's not dishonest by itself, but it can obscure the real return-driving error rate, especially for fitted or size-critical garment categories where a one-size miss still results in a return.
How long should I run a test before trusting accuracy numbers on my own store?
A holdout test of 60-90 days is generally enough to gather statistically meaningful data across your main categories, assuming reasonable traffic volume. Shorter tests are vulnerable to seasonal noise, promotional spikes, and small-sample variance, so treat anything under 30 days as directional only.
Does size recommender accuracy matter more than return-rate reduction?
No — accuracy is a proxy metric, and return-rate reduction is the outcome that affects your margin. A recommender with modest accuracy that meaningfully cuts size-related returns is more valuable than one with an impressive accuracy score that doesn't move your actual return numbers.