Helping shoppers find the right size

The final design is being prepared for release

Context

Returns caused by a poor fit are expensive: the product travels to the customer and back, inventory is tied up, and the customer is left without the item. My team’s entire workstream therefore serves one business goal: helping shoppers choose the right size the first time.

We already had AI size recommendations. A complex algorithm analyzed purchase history, brand size charts, and the experiences of people with similar preferences, with ongoing improvements from the ML team. Technically, it considered a lot. But research consistently showed a lack of trust. Users believed it ignored brand sizing, fit, and their body shape: exactly the factors the algorithm actually considered.

In parallel, I worked on a related task: redesigning the size guide. This is the reference shoppers can check for themselves before trusting a recommendation. Both projects grew out of the same problem and research, so I treated them as one sizing initiative with two complementary approaches.

Problem

We drew on the team’s quantitative and qualitative research to identify the problem. These were the main findings.

Choosing a size is genuinely difficult

In a survey with 3,100 responses, we asked what made choosing a size difficult. 52% were confused by unfamiliar brands’ size charts, 33% said charts did not account for body shape, and 27% reported errors or inaccuracies in size tables.

Shoppers do not trust size recommendations

When asked how the system knew a size would fit them, people mentioned purchase history and body measurements. Almost nobody mentioned comparing brand size charts or the algorithm’s work. Its most valuable contribution was invisible, making the recommendation feel superficial.

Most shoppers know their measurements approximately

65% knew their measurements; only 35% did not know them at all. People were comfortable with a margin of 1–3 cm and often chose a size based on one key measurement, such as the waist. This matters because recommendations based on measurements are about comparing choices made by similar shoppers, not simply matching numbers to a chart.

Hypothesis

If we make the AI’s work visible and explain what it considered — body measurements, previous purchases, and the specific brand’s size chart — while making the recommendation easier to notice, then shoppers will trust it more, use it more often, and choose sizes more accurately.

Solution

I split the work into two connected parts and developed them in parallel.

The first was a redesigned size guide: a reference shoppers could check and trust themselves. This part has already launched.

The second was an explainable recommendation: a redesigned prompt and a new flow that lets shoppers improve recommendations by adding body measurements. This was the core of the AI trust challenge. The concept went through several iterations and quantitative studies and is now being prepared for A/B testing.

Size guide

Before improving the recommendation, I strengthened the reference shoppers already trusted: the size chart. Research showed that it played an important role, with shoppers comparing their measurements against the table while choosing between two or three sizes.

I refreshed the visual design and increased the font size to make measurements easier to read. I added the model’s height, bust, waist, and hips to support the common habit of comparing oneself with the model. I also brought size advice into the guide and added clear measurement instructions. I generated the illustrations in Nano Banana to create a consistent set quickly, without involving an illustrator.

The experiment has been rolled out to 100% of users. Overall results were inconclusive, but in the iPhone segment, returns (ib_Return_rate) fell by 3% and conversion to opening the size chart rose by 0.84%. Our interpretation was that the iPhone audience, with greater purchasing power, was choosing more carefully: checking the chart more often and returning fewer items.

Size recommendation

I started with a hypothesis: displaying the model’s confidence percentage would increase trust. I tested it in quantitative research against the existing recommendation and a version without percentages.

Option 1: model confidence. The prompt showed a percentage, with progress bars for two sizes and an explanation of the factors considered on the detail screen.

Option 2: explainability, without percentages.

Option 3: the existing approach, showing information from reviews and listing the purchases used in the calculation.

Option 2, without percentages, won. This overturned my initial hypothesis. A confidence score did not increase trust and sometimes prompted the question, ‘Why isn’t it 100%?’ Trust grew when the system clearly explained what it had done. This matched the survey findings: people could not see the algorithm’s work, so we needed to explain its actions rather than assign them a number.

I refined the winning version by removing visual noise, highlighting the recommended size in the selector, and adding a clear entry point: ‘44 RUS should fit you · Why?’ Inside, three sections explain what we considered: your measurements, sizes from previous purchases, and the brand’s size chart. Each addresses a specific concern raised in research.

Ideas we ruled out

The solution took several iterations. We deliberately ruled out an explanation directly at the entry point because it overloaded the page; a long text explanation because people did not read it; showing only one recommended size because shoppers wanted to decide for themselves; and a generic explanation that did not connect to specific factors.

Collecting body measurements

After purchase history, body measurements were the input shoppers most expected the system to use. Size selection was also a natural place to ask for them. I designed a new flow in which shoppers add their measurements, helping the system match them with similar people and refine the recommendation.

The main risk was making the form intimidating. People remember measurements with some uncertainty, and half know them only approximately. I kept the form approachable: only measurements relevant to the category are required, marked with an asterisk — waist and hips for bottoms, for example. Everything else is optional. I also worked through keyboard behavior and movement between inputs.

I added a place in the profile to store and edit measurements, so shoppers always know where to find and update them.

Results

The size guide has launched to 100% of the audience. In the iPhone segment, returns fell by 3% and conversion to opening the size chart rose by 0.84%.

The new recommendation concept has been validated and approved. I took it through several iterations and quantitative user studies to check whether shoppers noticed the prompt and understood the basis of the recommendation. The winning version has been handed over to engineering and is being prepared for A/B testing, targeting lower returns (RRR) and higher CR to CartAdd. The measurement collection flow and its profile controls are being developed in parallel.