Sizing Advice Statistics for Conversion, Returns and Multi-Size Orders https://basketrescue.com/guides/sizing-advice-statistics-for-conversion-returns-and-multi-size-orders/ ## Quick answer A sizing recommendation can improve purchasing without reducing size-related returns. In the SizeFlags paper, a personalized-advice baseline increased conversion by 2.1% while its size-return reduction was below 0.5% and statistically insignificant. Separate 2017 tests of product-level size flags reduced size-related returns by 3.8% for shoes and by 4.3% or 6.6% for textile items, depending on the flag. These are relative changes from historical fashion experiments (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). For a merchant answering a sizing question, the important distinction is between helping someone buy and helping them buy the right size. This analysis follows the reported conversion, return and multi-size ordering outcomes through their separate experiments. It then sets out the product-level evidence a reply needs and the return window a merchant must preserve when evaluating the result. ## Key Takeaways The personalized-advice baseline improved conversion without a proven return reduction The SizeFlags paper reports a 2.1% relative conversion increase for its personalized-advice baseline, alongside a size-related return reduction below 0.5% that was not statistically significant (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). Product-level shoe flags reduced size-related returns by 3.8% A 2017 shoe experiment with 720,000 customers in each arm reported a 3.8% relative reduction in size-related returns under product-level size advice (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). Textile flag outcomes differed by direction The reported textile size-return reductions were 4.3% for too-small flags and 6.6% for too-big flags, a 2.3-point gap between the relative reductions (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). Sizing advice also changed multi-size ordering The shoe experiment reported size-selection order increases of 11.1% under too-small advice and 19.0% under too-big advice (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). ## Can sizing advice increase purchases without reducing returns? A shopper asks whether a jacket runs small. A confident answer may help them place the order. It may also encourage them to order two sizes, or steer them toward a size that fits. Those are different outcomes, so a single conversion number cannot tell a merchant whether the answer worked well. Figure description: Sizing advice statistics: applying the reported 3.8% relative return reduction to assumed starting rates of 10%,20% and 40% yields 3.8,7.6 and 15.2 avoided size returns per 1,000 orders. Figure 1: The same relative reduction can mean very different workloads. Source: SizeFlags, shoe-test relative reduction (https://arxiv.org/pdf/2106.03532), Jun 2021. Original analysis of the cited source. Source periods and calculation limits appear in the text. The SizeFlags research makes that distinction unusually clear. Its baseline personalized advice improved conversion, additions to cart and revenue per visit. The same baseline did not produce a statistically significant reduction in size-related returns. Buying became easier, but the evidence did not establish that the fitting problem had been solved. That matters when a team evaluates a shopping assistant. Completed orders show that a reply was persuasive enough to move a decision. They do not show that the selected size was right. Keep both outcomes in the evaluation, and allow the return window to mature before deciding what the advice achieved. ## How do personalized advice and product size flags differ? The baseline algorithm recommended a size using a customer’s order and return history. A product-level flag answers another question, whether this particular item tends to run too small or too big. A shopper with no purchase history can still use a reliable product flag, while a personalized recommendation depends on knowing something about that shopper. In the shoe experiment, each randomized arm included 720,000 customers. The size-flag intervention reduced the size-related return rate by 3.8% relative to its comparison group. The textile experiment used more than 180,000 customers per arm and reported separate reductions of 4.3% for too-small flags and 6.6% for too-big flags (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). These experiments help distinguish mechanisms, but they are not a head-to-head ranking of vendors or advice systems. The product groups, treatments and outcome definitions differ. A merchant should preserve those differences when choosing which kind of evidence to put into a sizing reply. ## Why can size warnings increase orders of several sizes? The shoe test also reported changes in size-selection orders, where a customer orders two or more sizes of the same item. Too-small advice increased these orders by 11.1%, while too-big advice increased them by 19.0%. The gap is 7.9 percentage points between the reported relative increases (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). This is an important result for the economics of recovered orders. A warning might prevent a shopper from abandoning, while also making a multi-size order feel safer. The merchant then has more items to ship and potentially more items to take back. Counting every extra order as equally valuable would miss that trade-off. If the catalog supports a clear alternative size, the reply can explain the specific evidence for choosing it. If the evidence only says that fit is uncertain, forcing a precise recommendation creates confidence the record does not support. Keep the shopper’s actual decision visible instead of optimizing the tone of reassurance in isolation. ## Where should the sizing evidence appear? A generic size chart gives measurements or conversion rules. A product flag concerns observed fit for one item or a supported group of related items. A customer’s previous purchases concern their own history. Those records should be identified separately before they are combined in a reply. For a recovered cart, start with the exact item and selected size. Check that the chart or fit record belongs to that version. Then explain what the evidence supports in plain language, such as the measurement difference between two sizes or the reason a product-level warning applies. Where the store lacks a useful fit record, ask for the measurement that would change the answer. The merchant can review a small set of actual size questions before rollout. Look for replies that silently switch from a product measurement to a universal claim about body shape. That is a factual problem even when the wording sounds helpful. ## How should the order be measured through its return window? Use separate measures for purchase conversion, items per order, size-selection orders and size-related returns. Record which advice the shopper received and whether it came from a personalized recommendation, a product flag or a general chart. Otherwise a changing mix of advice can make the result hard to explain. Compare groups over compatible periods and give returns enough time to appear. An early lift in purchases can arrive weeks before its fulfillment costs are known. Keep total returns separate from size-related returns, because the study’s sizing outcome excludes other reasons for sending an item back. For a randomized merchant test, report both the change in order completion and the later cost outcome. A small conversion improvement with no fitting improvement may still be worthwhile. Its value needs to be calculated from the actual orders, margins and returns of that merchant. ## Which findings can be cited with their original scope? The 3.8%, 4.3% and 6.6% figures are relative reductions in the reported size-related return rates. They are not reductions of that many percentage points in an underlying rate. The source does not provide a shared baseline that would justify converting all three into one common number of returns saved (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). The paper also describes later observational monitoring and production improvements. Those sections are valuable for understanding the system, but they should not be merged into the randomized results. A citation should name the intervention, product category, historical period and outcome being quoted. The practical lesson is to evaluate the complete sizing decision. Ask whether the shopper bought, which sizes they ordered and what happened after delivery. That gives a merchant a more useful assessment than a single reassuring-answer score. ## Which reported measures support these findings? The figures below preserve the source’s outcome labels and units. Read comparisons within their stated experiment, population and period. Reported measure | Value and unit | Source Personalized advice conversion change | 2.1 percent relative increase | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Personalized advice add-to-cart change | 1.8 percent relative increase | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Personalized advice revenue-per-visit change | 2.1 percent relative increase | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Upper bound on the nonsignificant size-return reduction | <0.5 % | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Shoe SizeFlags size-return reduction | 3.8 percent relative reduction | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Textile too-small flag return reduction | 4.3 percent relative reduction | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Textile too-big flag return reduction | 6.6 percent relative reduction | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Too-small shoe selection-order change | 11.1 percent relative increase | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Too-big shoe selection-order change | 19 percent relative increase | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) Customers per shoe-test arm | 720000 customers | (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021) ### How are the additional calculations reproduced? These are arithmetic comparisons of the published inputs. They retain the same sample and limitations as the inputs and do not create an additional experiment. Calculation | Expression | Result and unit Gap between textile flag relative reductions | 6.6 - 4.3 | 2.3 percentage points Gap between selection-order relative increases | 19.0 - 11.1 | 7.9 percentage points Relative size of too-big versus too-small selection increase | 19.0 / 11.1 | 1.7117 ratio Reported shoe-test arm exposures | 720000 * 2 | 1440000 customer assignments ## How many returns does a relative reduction actually avoid? A 3.8% relative reduction has very different operational size at different starting rates. For 1,000 otherwise comparable orders, an assumed 10% size-return rate means 100 size returns; applying a 3.8% reduction gives 96.2, or 3.8 avoided returns. At a 20% starting rate, the corresponding saving is 7.6 returns. At 40%, it is 15.2. These are worked sensitivities using the reported shoe-test relative reduction, not reconstructed counts from the experiment. (SizeFlags researchers, relative-effect input (https://arxiv.org/pdf/2106.03532), Jun 2021). Assumed starting size-return rate | Rate after a 3.8% relative reduction | Avoided size returns per 1,000 orders 10% | 9.62% | 3.8 20% | 19.24% | 7.6 40% | 38.48% | 15.2 If each avoided return saves an assumed $12 in handling and other variable costs, the corresponding savings are $45.60, $91.20 and $182.40 per 1,000 orders. That leaves very different room to pay for advice or extra fulfillment. The model excludes lost sales, retained item margin and other returns, so it is a narrow cost comparison, not total profit. ### Which denominator should follow an increase in multi-size orders? Measure the number of extra units shipped as well as the count of orders containing several sizes. One order containing two sizes and one containing four sizes both count as selection orders but create different outbound and return workloads. The source's 11.1% and 19.0% changes concern selection-order incidence; they do not give a common average extra-unit count. Nor should the 19.0% too-big subgroup response be combined with the overall 3.8% shoe result as if both described one treatment population. (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). A merchant can reconcile the economics using its own joint records: contribution from retained merchandise, less extra outbound cost, reverse-logistics cost, write-offs and advice cost. Preserve the same assigned-customer denominator for the comparison. Recalculating only among purchasers conditions on an outcome the advice may itself change. ### What result would distinguish reassurance from better fit? A useful diagnosis follows purchase, size-selection behavior and matured returns together. Higher conversion with unchanged fit returns is consistent with reassurance or reduced hesitation. Lower size returns with unchanged multi-size ordering is more consistent with improved fit selection. Higher multi-size ordering alongside lower size-related returns requires the extra-unit costs before a commercial conclusion. These are competing interpretations to investigate, not causal mechanisms proven by the public statistics. Test a familiar-size repeat purchase as a transfer case and a new product with no personal history as another. The personalized baseline and product-level flags solve different information problems. Keep a general chart available, verify the exact item/version and avoid implying that a product flag identifies a person's body measurements. The baseline-rate sensitivity (https://basketrescue.com/guides/sizing-advice-statistics-for-conversion-returns-and-multi-size-orders/data/size-return-sensitivity.csv) and calculation (https://basketrescue.com/guides/sizing-advice-statistics-for-conversion-returns-and-multi-size-orders/analysis.py) make the arithmetic reusable with a store's real starting rate and cost. The public paper provides a relative effect in a historical population; the merchant supplies the denominator and economics needed for its decision. ## Methodology and Sources The comparison transcribes the separate interventions in sections 4.2 and 4.3 of the 2021 SizeFlags paper. Baseline personalized recommendations, product-level too-small/too-big flags and observational monitoring are treated as separate evidence. Calculations compare published relative changes and the explicitly stated shoe-test arm sizes. No common starting return rate is assumed and no results are projected onto a current merchant (SizeFlags researchers (https://arxiv.org/pdf/2106.03532), Jun 2021). SizeFlags researchers: June 2021 paper. Citation months identify the cited publication or version. Where a citation explicitly says “source checked”, November 2026 is the verification month, and the original study period remains stated in the text. Source observations, calculations and the proposed merchant applications have distinct meanings. All evidence is historical Zalando fashion research. Relative reductions are not percentage-point changes in the underlying return rate. The baseline size-return reduction was statistically insignificant and reported only as below this bound. The paper reports 720 k in each arm. Other experiments have separate samples. ## Glossary Relative change: A change divided by the original level. Conversion: The purchase outcome defined by the experiment. Size-related return: A returned purchase classified as a size or fit problem. Selection order: An order containing alternatives, such as several sizes. Control group: The comparison group that did not receive the tested intervention. ## Sources SizeFlags researchers: SizeFlags: Reducing Size and Fit Related Returns in Fashion E-Commerce (https://arxiv.org/pdf/2106.03532) Check the source figures The data file retains source labels, values, populations and limitations for this analysis.