A Fit Pilot Should Earn Its Rollout: A Commercial Scorecard for Footwear Teams
A footwear fit pilot can produce a busy dashboard and still leave the commercial team with an unanswered question: should we roll it out?
Recommendation starts, completed scans and clicks tell you whether people use the experience. They do not, by themselves, tell you whether the business is better off. The decision needs to connect customer behavior with the orders people keep, the cost of serving them and the effort required to operate the service.
Here is the evaluation framework we recommend for footwear teams considering AI Fit Intelligence. It is a proposed decision framework, not a claim that every Perfitt deployment produces the same outcome.
AI-generated editorial illustration, not an actual customer deployment.
Start with the decision the pilot must support
Before launch, write down the next decision: expand to more footwear models, add a market, continue testing, or stop. Then define the eligible assortment, audience, comparison method and observation period. A pilot covering one collection should not silently become evidence for an entire global catalog.
Choose products with enough traffic, usable stock and reliable order and return records. A recommendation cannot rescue a size that is unavailable. Keep a record of stock gaps, promotions, price changes and acquisition campaigns so the team can interpret the result in context.
Commercial, ecommerce, customer service and analytics owners should agree on the scorecard together. That agreement prevents a familiar problem: marketing reports more purchases while operations is still waiting to see what comes back.
Build a scorecard that separates adoption from value
Customer reach and experience
Track how many eligible shoppers can access the recommendation, how many start and complete it, and where they leave. Report coverage alongside usage: a low completion count means something different when the service is visible on only a small fraction of eligible product pages.
Also inspect errors, mobile usability and requests for help. These are diagnostic measures. They show where an experience needs attention before you interpret a weak commercial result as a verdict on fit guidance itself.
Purchases that survive the return window
Measure purchase conversion, then revisit the same order cohorts after their return window has matured. Track retained revenue, all returns, size-related returns and exchanges separately. Report how many return reasons are missing; an improvement in a category with incomplete coding deserves caution.
Define each denominator. A return rate per item is not interchangeable with a return rate per order. An exchange is not necessarily lost revenue, but it can still create shipping and handling costs. Keep these distinctions visible.
Contribution after the relevant costs
For the commercial decision, compare retained sales less the costs the retailer actually incurs: cost of goods, fulfillment, reverse logistics, payment costs and attributable support costs. Account for software fees, implementation and ongoing operating effort as well.
Use contribution per eligible visitor or shopper on a consistently defined basis when the pilot is intended to change conversion. Comparing only successful orders can miss people who would not otherwise have purchased. Agree with finance on where each cost belongs so a return-handling charge is not subtracted twice.
Five definitions to agree before the first report
Coverage: covered eligible models divided by eligible models; report shopper exposure separately.
Engagement: starts per eligible exposed shopper and completions per start. Use these to diagnose the journey.
Retained conversion: assigned eligible shoppers who retain at least one item divided by assigned eligible shoppers. Retained net revenue is merchandise sales less discounts and refunds, on a consistent tax basis.
Returns: returned units divided by fulfilled units after equal return-window maturity. Separate size-related reasons, unknown reasons and exchanges, then examine model-size cohorts.
Contribution: retained net revenue less retained-unit cost of goods, fulfillment, reverse logistics, payment and attributable support costs. Reconcile recovered inventory and write-downs. Count each cost once, and account for technology and implementation costs over an agreed horizon.
AI-generated editorial illustration, not an actual customer deployment.
A fair comparison matters more than an impressive percentage
People who choose to use a fit tool may differ from people who ignore it. A higher purchase rate among users can be an interesting signal without establishing that the tool caused the difference.
Where feasible, use a randomized comparison with stable assignment and agreed eligibility. Compare the groups as assigned, then use engagement data to understand the result. Microsoft's researchers discuss the principles and pitfalls of web experimentation in their Practical Guide to Controlled Experiments on the Web.
If randomization is unavailable, document the limitations of a before-and-after or market comparison. Seasonality, stock availability and traffic quality can move the same metrics. Present observational findings as observational findings.
Set the evaluation period and analytical decision rules before reviewing outcomes. Ask the analytics owner to size the test for a commercially meaningful effect and report uncertainty. When the evidence is inconclusive, say so; a point estimate alone should not decide an expansion.
Make the rollout decision explicit
Expand when the agreed commercial objective is met with adequate evidence, customer experience remains acceptable and the operating team can support broader coverage.
Repair and retest when a specific issue—such as limited coverage, recommendation errors or an unclear entry point—prevents a useful assessment. Record the change and the new test scope.
Hold or stop when costs outweigh the demonstrated benefit, important safeguards deteriorate, or the evidence remains too weak to justify expansion.
Do not use a pilot's end date as a substitute for a decision rule. Return data may lag checkout data, and a small sample may remain uncertain even when the calendar says the test is finished.
Agree the data handoff and review cadence
The retailer and implementation team should agree how eligible exposure, recommendation interactions, orders, refunds and exchanges can be joined using permitted identifiers. Record consent requirements, data retention, access controls, missing return reasons and failed joins. This scorecard proposes what to measure; it does not imply that every metric is available automatically within Perfitt.
An optional 30/60/90-day plan can help organize the work: define and instrument, observe and diagnose, then review matured outcomes. These are planning checkpoints, not a guaranteed rollout timeline. Traffic and return policies determine when a useful decision is possible.
What this means for AI Fit Intelligence
PerfittSize supports product-specific size guidance, while PerfittOnsite brings measurement and recommendation into offline experiences. The measurement framework above must be scoped with each retailer. The immediate question is whether a better match between a person and a product improves the shopping outcome. Over time, fit and demand data may also help teams examine assortment and size-level inventory decisions. Those broader opportunities deserve their own validation; a conversion result does not automatically prove a merchandising or production benefit.
The same discipline applies to Zero Waste ambitions. Fewer returned pairs may be a useful operational signal, but it is not an emissions calculation. Environmental claims need an appropriate measurement boundary and evidence.
A useful pilot leaves the team with a decision it can defend: what changed, for whom, at what cost, and what should happen next.
Planning a footwear fit pilot? Bring your target assortment, current return definitions and commercial objectives to a 15-minute Fit Pilot Discussion with Perfitt. Start with the business question your team needs answered.