AI Size Recommendation for a Clothing Brand From Charts and Returns

A small clothing brand can build an AI size recommendation by turning its size chart into a table of garment measurements, adding fit notes for each style, coding past return reasons, and giving all three to an assistant with a rule to say "not sure" rather than guess. Start with a fit guide and a short question form.

The data does the work, not the tool. A recommender is only as good as the garment measurements and the return notes behind it, and most small brands already hold both in scattered places: a PDF size chart, a supplier spec sheet, and a returns inbox. The steps below pull them together, show how to check the recommendations against real returns, and compare the two ways of putting this in front of customers, a recommender widget and a chatbot.

Follow me on Instagram@sagnikteaches

The example brand and its problem

Take an illustrative brand selling women's and men's casual wear online through Shopify: 1,200 orders a month, about 20 per cent returned, so roughly 240 returns. The owner exports the return reasons from the returns portal and codes them into five groups. This is the starting picture, and it is made up for the illustration, but the method works on any numbers.

Connect on LinkedInSagnik Bhattacharya

Size-related returns total 156 a month, 65 per cent of all returns. If each return costs the brand about $9 in postage, handling and lost resale time, that is roughly $1,400 a month in size returns alone. The target for this exercise: cut size returns by around a fifth, which saves about $280 a month and, more importantly, fewer disappointed customers. These figures are illustrative; substitute your own cost per return from your carrier invoices and staff time.

Subscribe on YouTube@sagnikteaches

Step one: build the garment measurement table

A size chart usually lists body measurements for sizes S to XL. A recommender needs two tables: the body measurements each size is designed for, and the actual garment measurements for each style, which differ because of ease, stretch and cut. Ask the supplier for the spec sheet, or measure a sample of each size flat with a tape. Record in centimetres, and note the fabric stretch.

The two chest columns are the important lesson. A customer with a 94 cm chest sits in size M by body measurement, and a relaxed-fit tee with a 102 cm garment chest gives them 8 cm of ease. For a slim-fit shirt of the same brand, the garment chest might be 98 cm, giving 4 cm of ease. Put one table per product family in a sheet, and give each style a "fit type" label: slim, regular, relaxed or oversized. If you skip the garment table and give the AI only the body chart, it will recommend the same size for every style, which is exactly why customers return things.

Step two: write fit notes in plain sentences

Measurements miss what staff know from handling the garments. Spend an hour with whoever packs orders and write one or two sentences per style, in a column next to the table. Illustrative examples:

• "Runs small across the shoulders; size up if broad-shouldered or between sizes."
• "Fabric has 5 per cent elastane and gives about 3 cm. True to size on the chest, short in the body for people over 183 cm."
• "Sleeves are long. Customers under 165 cm often say they need to roll them."
• "Oversized cut: most customers take their usual size for the intended look and one size down for a closer fit."

These notes are the cheapest improvement in the whole project, and a good customer reads them too. They also belong in your product description; writing Shopify product descriptions with AI shows how to put a fit line near the top without a long essay.

Step three: turn return reasons into per-style rules

Export returns for the last six to twelve months. For each return you want the style, the size ordered, the reason, and any free-text comment. Code the reasons into the groups above; an assistant can help with the coding. The prompt below works for batches of 50 comments.

Below are customer return comments from a clothing shop. For each,
give: STYLE (as written), SIZE ORDERED, REASON CODE from this list:
TOO_SMALL, TOO_LARGE, TOO_SHORT, TOO_LONG, TIGHT_SHOULDERS,
TIGHT_HIPS, STYLE_COLOUR, QUALITY, OTHER, UNCLEAR.
If the comment does not say, use UNCLEAR. Do not guess.
Return a table. Do not add information that is not in the comment.

COMMENTS: [paste]

Check a sample of 20 yourself against the original comments, since a few will be mis-coded. Then count by style. In an illustrative result, the canvas overshirt has 18 returns in a month, 14 coded TOO_SMALL, nearly all from customers who ordered their usual size. That is a clear instruction: add "runs small; size up" to the fit note and the recommender rule, and tell the supplier if the pattern holds. A style with scattered reasons and no clear direction needs no rule at all.

The bigger programme of reading complaints and returns for patterns, not only sizing, is in spotting patterns in complaints, returns and defects with AI. The sizing exercise here is a focused use of the same idea.

Step four: write the recommendation rules and test the logic

Now you have three inputs: the body chart, the garment table with fit types, and the per-style notes from returns. The recommendation logic can be written as short rules, which an assistant follows or which you set up in an app:

1. Ask for height and usual size, and optionally chest or waist, and the fit the customer wants (close, regular, roomy).
2. Find the size whose body range contains their measurement. If they give none, start from their usual size in other brands.
3. Adjust for the style's fit type: slim styles, size up if the customer wants room or sits at the top of a range; oversized styles, offer the usual size and one down.
4. Apply the style's note from returns, such as "runs small, add one size".
5. If the customer is between two sizes or any input is missing, say so and give both options with the reason.

Test it on 20 past orders where you know the outcome. Take 10 kept orders and 10 returned for size, give the recommender the customer's details as best you can reconstruct them (height from the order note, or from the customer if they agree to be contacted), and see whether it would have suggested the size they bought. For the 10 size returns, would it have recommended a different size? Even a rough test of this kind finds nonsense rules before customers do.

Step five: choose a widget or a chatbot

There are two ways to put the logic in front of customers, and they suit different brands.

For most small brands the sensible order is: first a fit-guide table and fit notes on the page, costing nothing; second a widget that asks the three questions; third a chatbot only if message volume about fit stays high. The Shopify App Store has dedicated size-recommendation apps, and each lists its own pricing and what it needs from you, so compare the current listings and any free trial before paying. Read the data requirements and the billing carefully: one popular listing, for example, imports size charts from images, product descriptions or a CSV, caps the number of charts and recommenders by plan, and adds charges on its higher plans as your order volume grows.

If you do want a chatbot, the best AI chatbots for small Shopify stores compares options, and whether an AI chatbot can handle order tracking and returns covers the neighbouring jobs a sizing bot is often asked about. A plain chat assistant such as ChatGPT or Claude can also be used behind the scenes: load your tables and notes into a Project and have staff paste in customer questions to get a drafted answer, a good middle step before anything customer-facing.

A prompt for the staff-assisted version

You help a clothing shop recommend sizes. Use ONLY the size tables,
fit types and style notes below. If the customer's details do not
let you choose one size confidently, give two sizes and explain the
difference in one sentence each. Never invent measurements. Keep
replies under 80 words. If a style has no note, say so.

[paste tables and notes]

CUSTOMER: 168 cm, usually a medium in high-street shops, prefers a
roomy fit. Asking about the canvas overshirt.

An illustrative reply: "For the canvas overshirt I'd suggest size L. It runs small across the shoulders, and you want a roomy fit, so one size above your usual medium gives you the room. If you'd prefer a neater look, M is the alternative, but it will be snug over a jumper." That is the behaviour to want: a clear pick, the reason drawn from your note, and an honest alternative.

Where size recommendations go wrong

Body charts treated as garment charts

The most common failure. A recommendation that ignores ease and fit type recommends the same size across a slim shirt and an oversized hoodie. Always include the garment table and the fit type.

Thin return data leading to false rules

Four returns of one style, three for "too small", looks like a pattern and may not be. Set a minimum before you write a rule: for example, at least 10 returns or 10 per cent of that style's orders. Below that, mark the note "watch" and look again next month.

Confident answers outside the data

A chat assistant asked about a style that is not in your table may quietly produce an answer from general knowledge. Include a rule that a style not found in the table gets "I don't have fit data for this style, please ask our team". Test it with a made-up style name.

Sensitive measurements collected carelessly

Height and a usual size are enough for most recommendations. If you ask for body measurements, say why, store as little as possible and check your privacy notice covers the use. Do not paste customer names and addresses into a chat assistant to draft a fit reply; paste the measurements and the style only.

Different batches, different fit

Suppliers change patterns. If returns for a style that used to be fine suddenly spike after a new delivery, measure the new stock against the old table before blaming the recommender.

Three customers walked through the rules

Rules only prove themselves on actual cases. Here are three illustrative customers run through the five steps above, using the measurement table, a slim-fit shirt and the canvas overshirt from step three.

Customer one: clear measurements, a slim-fit shirt

A customer gives a 94 cm chest and wants a close fit. Body range puts them in M (90-95). The slim shirt's garment chest is 98 cm, so M gives 4 cm of ease, right for a close fit. No style note applies, so the recommendation is M, with the reason stated: "M gives you about 4 cm of room across the chest, a neat fit." Simple, and the widget handles this case perfectly.

Customer two: between sizes and wanting room

A customer is 168 cm, gives no measurements, says they are "usually a medium", and wants a roomy fit in the canvas overshirt. Step two starts from their usual M. Step three leaves it alone because the overshirt is regular fit. Step four applies the note "runs small across the shoulders, size up if wanting room", so the answer is L, with M as the alternative for a neater fit. A fixed widget that skips the style note would have said M, and this customer would have been one of the 14 returns.

Customer three: outside the data

A customer asks about a style that is not yet in the table, a new knitted vest. The rule says "no fit data for this style" and passes the question to a person, who replies with measurements from the sample. That is the correct outcome; a recommender that answers anyway is the one to be afraid of. Note the question in the sheet as a reminder to add the vest before more customers ask.

What to do with returns that still come back

No recommender removes size returns, so make the remaining ones teach you something. When a customer starts a return coded as a size problem, reply with a short message that asks one question and offers an exchange. A template with placeholders:

Hi [first name], sorry the [style] didn't fit. To help us get the
right size next time, could you tell us your height and which size
you ordered, and whether it was too tight, too loose or too short?
We can send the [suggested size] as an exchange straight away.
Thanks, [your name]

The reply gives you a structured answer for the sheet and often turns a refund into an exchange. In the illustrative numbers, if one in four size returns becomes an exchange, the 156 size returns a month produce 39 exchanges: 39 sales kept rather than refunded, although each exchange still carries its own postage and handling cost. The rules for refunds and exchanges, including when a person must approve, are covered in automating returns and refunds with clear AI rules.

Measuring size returns before and after

Run it for eight weeks and compare against the eight weeks before. Track the number of size-coded returns as a share of orders for the styles covered, the number of fit questions arriving by email or chat, and conversion on those product pages if your analytics allow. In the illustrative example, 96 "too small" and 60 "too large" returns across 1,200 orders is a size return rate of 13 per cent. A fall to around 10.4 per cent of orders, which is the one-fifth cut, would mean about 31 fewer returns a month.

Watch for a side effect: if "too large" returns rise as "too small" falls, the rules are overcorrecting by telling everyone to size up. Look at the split by style and fix the notes for the ones that moved the wrong way. Keep a note each month of what you changed, so a change in the numbers can be traced to a change in the rules.

Keeping it alive after launch

Each month, add the new return reasons to the sheet, recode the style counts, and update any note where the pattern has moved. When you launch a new style, write its garment table and fit note before it goes live, not after the first wave of returns. When the supplier changes, remeasure. The routine takes about an hour a month and is the reason the recommendation stays right.

If you sell other kinds of goods the pattern carries over. A sports club shop selling kit has team-specific cuts; a pet shop selling coats and harnesses has measurements by neck, chest and back length; a gift shop selling jewellery has ring sizes and chain lengths. In each, a measurement table, a fit note from the people who handle the goods and a coded set of return reasons feed a short set of rules. If you would like help setting this up for your own range, that is within my AI implementation consultation, via the contact page. Start with the return reasons: they are free, they are yours, and they tell you where to begin.

Questions about AI sizing for a small brand

Can a chat assistant recommend sizes accurately?

It can apply a size chart and fit notes you supply, which is reliable for simple rules. It cannot know how a garment fits without that data, and it can guess confidently when the data is thin. Give it measurements and a clear instruction to say when it is unsure.

How much returns data do I need before this is worth doing?

A few hundred orders with return reasons is enough to see which products run small or large. Below that, rely on your garment measurements and fit model notes. With very small volumes, one or two odd returns can mislead you, so look at patterns across styles rather than single items.

Should I use a size widget app or build a chatbot?

If most buyers want a quick answer on the product page, a widget that asks two or three questions suits better. A chatbot suits brands with complicated fits and a lot of questions by message. Many small brands start with a plain fit-guide table and a short question form before paying for either.

Do I need customers' body measurements?

Not always. Asking height, usual size in another brand and fit preference often gives a good recommendation without sensitive measurements. If you do collect measurements, say why, store as little as you can and check your privacy policy covers it.

Further reads

• How Much Does It Cost to Add AI to a Shopify Store? — What adding AI to a store costs in fees and time.
• Your First 30 Days of AI Customer Support for an Online Shop — A month-by-month start for AI customer support.
• Can a Small Online Shop Use AI to Compete With Bigger Brands? — Where a small shop can use AI against bigger brands.
• How to Set Up AI Abandoned Cart Emails for a Small Online Store — Recover carts lost to sizing doubt with timely emails.
• How to Create Outfit and Styling Posts for a Boutique With AI — Show fit and styling on real bodies in social posts.
• AI Fit Shoes Plugin: Can a Size Finder Cut Returns for a Shoe Shop? — How shoe size finders work, what data they need, and a before-and-after method to test whether one cuts your returns before you pay.
• AI Mirror for a Clothing Store: Is a Smart Fitting Room Worth It? — A small-store payback sum for an AI smart mirror, the vendor claims to test, camera privacy questions and cheaper ways to get the benefit.
• AI Tools for Clothing Brands: From Sketch to Tech Pack to Listing — One overshirt taken from sketch photo to tech pack, supplier brief and product page, with the prompts, numbers and checks for each handover.
• Clothing Return Policy Template: Write One With AI and Check It — A decision sheet, an AI drafting prompt that audits itself, fourteen legal checks and twelve test scenarios for a clothing shop's return policy.
• Can a Chatbot for a Clothing Store Answer Size and Fit Questions? — The size chart as source, a six-step answer flow, two worked chats, return-policy limits, hand-over triggers and 30 test questions for a clothing store.
• Ecommerce Product Recommendation AI: Quiz, Widget or Chat Assistant? — A quiz, an on-page widget and a chat assistant costed for a small store, with a rule by catalogue size and a method to measure add-to-cart lift.
• AI Marketing for a Clothing Brand: One Drop Planned Start to Finish — Plan a complete clothing drop from teaser to restock using ChatGPT and Canva with exact inventory numbers and asset schedules.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Shopify App Store listing for Kiwi Size Chart & Recommender (features and billing), Claude pricing page (Projects on the Team plan) (checked 5 Nov 2026).