Size Advice in Chat: Avoiding Returns Up Front

A jumper cannot be tried on online. What a fitting room settles in two minutes in a physical store is decided in online retail only once the parcel has been opened, and the merchant then carries the consequences. Size and fit are therefore a frequent reason for returns, and unlike transport damage this reason can be removed before the purchase: the information almost always exists, just not where it is being read. It sits in the manufacturer's size chart, in the measurements of the variant, in the material note, sometimes with an item the same customer has already bought. A chat assistant can bring these details together at the decisive moment, but only under one condition: it has to read from the stored chart instead of from what sounds usual for a garment. This article shows which data are needed for that, how the comparison with an item already bought works, what to do when measurements are missing and where the line runs between a fit note and a promise.

Enlarge graphic

Labels in the illustration

• Shop chat on the product page
• I usually wear L. Does this fit?
• Size L measures 106 cm chest here.
• Per the catalogue this series runs narrow.
• With broad shoulders, rather take XL.
• Compare: your rain jacket in L?
• The rain jacket measures 110 cm in L.
• This model runs 4 cm narrower.
• For the same fit, XL works here.
• Read from the size chart
• Measurements given, no promise of fit
• Write a message
• What the recommendation is built on
• Size chart of the manufacturer per series
• Variant measurements: chest, waist, length
• Material, elastane content and cut
• Fit note as a maintained data field
• Comparison item named by the customer
• If a measurement is missing, the assistant says so.
• Two routes, one question
• Estimated
• L is usually a 40
• Two sizes ordered
• One of them goes back
• The case costs twice
• From the chart
• Measurement, not a rule
• Compared to an item owned
• The gap is named
• The answer holds up
• Settle the size before the purchase instead of processing the return after it

Size advice in chat before the purchase

Why size only becomes apparent after delivery

With almost every other product attribute the customer can check before buying whether it is right. Colour, features, connectors and scope of delivery are on the page and can be held against one's own needs. Fit cannot: it only shows on the body, and therefore only after dispatch, delivery and unpacking. Several days and the entire cost chain of an order lie between the decision and its verification. That is exactly what makes size an expensive special case: the mistake is made in the cart, noticed in the living room and paid for in the returns warehouse.

Matters are made worse by the fact that the size label itself says little. A clothing size is not a binding standard but a house convention: the same label stands for different measurements at two manufacturers, across two series from the same house and in two countries. With shoes the conversion between length in centimetres and size system is added, with trousers the combination of waist and inside leg, with outerwear the cut. Anyone reading only the letters or numbers in the shop is comparing figures that are not comparable. Customers notice this and fall back on the method that always works for them: they order two sizes.

The manufacturer's size chart as the basis

For a size answer to hold, the assistant needs read access to the places where the shop keeps its measurements. As a rule these are the size chart per manufacturer and series, the variant measurements on the article, the material details including elastane content, and the cut. Added to that are details often overlooked in daily work: the model's measurements from the image caption, the care note on shrinkage and the manufacturer's conversion table between country sizes. For Shopware in the freely available Community Edition the assistant takes articles, variants and descriptions from the shop; size charts and measurements that are not there are stored as documents in the knowledge base. Which details an assistant can evaluate in a shop is set out on the page about the shop assistant; which actions are permitted is defined by tool control.

• Size chart per manufacturer and series, in the same version that appears on the product page
• Measurements of the individual variant: chest, waist, inside leg, sleeve length, insole length
• Material, elastane content and weave, because a stretchy fabric has a different tolerance than a firm one
• Cut and fit as a maintained field, for instance slim, regular or wide
• The model's measurements from the product image, so the visual impression gets a reference point
• Conversion between the manufacturer's country sizes instead of a generic table found on the web

Whatever cannot be read from the system belongs in a maintained knowledge base: known deviations of individual series, feedback from customer service, the experience that a certain model comes out shorter after the first wash. What matters is that this has an owner and a rhythm, otherwise such text ages faster than the catalogue. What such a maintenance process looks like is described in the article Building a knowledge base: which content belongs in. The separation is deliberate: measurements come from the catalogue, experience and exceptions from the knowledge base, and each is checked separately.

From the wish to the measurement: how the conversation leads

Customers do not describe their goal in centimetres. They say "I usually wear L", "I have broad shoulders" or "it should sit loosely". An assistant that derives a size from this straight away has guessed. The workable route runs through a reference point: either a body measurement the customer can take, or a garment that fits them and whose measurements are known. Only with such a reference point does the size chart turn into a statement about this one case. That sets the order: first establish the reference, then hold it against the chart, then justify the result with the measurement.

• Establish the reference – which measurement or which comparison garment is available?
• Ask how it should sit – should the item be close-fitting, straight, or worn over something?
• Hold it against the chart – compare the stated measurement with the stored values of the variant.
• Justify the result – state the size together with the measurement, not just the letter.
• Name the borderline case – if the measurement falls between two sizes, say so instead of smoothing it over.

Two or three follow-up questions are enough for this. More feels like an interrogation and costs exactly the patience that is scarcest just before a purchase. Part of that is asking only what the recommendation needs: a measurement in centimetres is enough for a size answer, body weight is no substitute for it and has no place in the conversation. How to build an advisory dialogue that narrows down instead of interrogating is shown in the article Guided selling in chat: finding the right item. If the customer can photograph a label or a tape measure, the reference point can also be captured that way; the article Image upload in AI chat describes what to watch out for.

The comparison with an item already bought

The most reliable size answer a shop can give is not a piece of advice but a juxtaposition: the item the customer already owns and wears is placed next to the item they are considering. Both sets of measurements are in the catalogue, and the difference is arithmetic. "Does L fit me?" thus becomes a sentence nobody has to interpret: "The jacket you mentioned measures 110 cm across the chest in size L; this model measures 106 cm in L and therefore runs four centimetres narrower." The customer then decides for themselves, and decides on a basis they know, because the comparison garment is hanging in their wardrobe.

Two things are required for this, and both have to come together: the customer names the article they own, and measurements for both articles are in the catalogue. The assistant does not access earlier orders, and an email address mentioned in the chat does not open them either. If the measurements on the older article are missing, because it was delisted or never fully maintained, the only route left is a measurement the customer takes themselves. How advice in chat leads to the right article is described in the article on guided selling in chat; the framework for processing and hosting is set out under privacy and hosting.

When measurements are missing: naming the gap

No catalogue is fully maintained. In practice there are three states: measurements exist per variant, only a chart at series level exists, or there is nothing. The assistant has to tell these three states apart, because only the first one supports a concrete recommendation. Where the basis is missing, the only workable answer is the one nobody likes to give: nothing is stored. In our own measurement series 8 out of 8 questions without a basis were declined instead of guessed (XICBOT, own measurement series of 11 September 2026); eight questions deliberately placed outside the available material were tested, and the wording of every answer is in the measurement file. The example question concerned contact details that were not stored; the behaviour with a measurement that is not stored is the same. It can be read up under measurements.

That does not end the conversation, though. Even without a measurement four routes remain open, and they make sense in this order: state the measurements of a comparable article from the same manufacturer, ask the customer for a measurement of their own, hand over to a member of staff who asks the manufacturer, or point to the right of withdrawal that applies anyway. Each of these routes is honest; an estimated figure in centimetres is not. How the binding to one's own material is created technically is described in the article Preventing hallucinations with a knowledge base and RAG.

Fit notes come from product data

"Runs small" sounds like a turn of phrase, but it is a statement about the goods. It should therefore be treated like any other product detail: as a field that somebody maintains, checks and answers for. In practice a small vocabulary is enough, for instance runs small, true to size, runs large, plus a free-text field for the reason. If a language model does not find this field, it fills the gap with a plausible phrasing, and with clothing almost anything is plausible. The difference between a reliable and an invented fit note is therefore not decided in the assistant but in the product data.

What applies legally when advice is given in chat

The chat is not a side room with rules of its own. What is said there about the goods is a statement about the goods. The pre-contractual information duties in distance selling, among them "the main characteristics of the goods", are set out in Article 246a Section 1 of the Introductory Act to the German Civil Code. For assessing incorrect statements, Section 5(2) no. 1 of the German Act Against Unfair Competition applies: statements about the main characteristics of the goods are expressly covered there, and untrue statements about them count as misleading. An invented fit note is therefore not a stylistic slip in a conversation log but a statement the merchant will be held to.

This right of withdrawal exists independently of any advice. A size answer in the chat shortens no period, replaces no withdrawal instruction and changes none of the merchant's duties; it only reduces the reason to make use of the right. Conversely: the conversation log can be read and stored, on both sides. What the assistant said about measurements and fit can be produced by a customer later. That is why every figure in the chat belongs in the same check as every figure on the product page. This description does not replace legal advice; how complaints and warranty are kept apart in a conversation is covered by the article Complaints and warranty in chat.

Settle it before the purchase instead of processing it after

A size question has two possible outcomes, and they cost very differently. If it stays open, it ends in abandonment or in a selection order with a return planned from the start. Across the average of fifty separate cart abandonment studies, 70.22 % of online shopping carts are abandoned (Baymard Institute, as of 22 September 2025); not every one of these abandonments is down to size, but the size question is one of the few a conversation can answer in seconds. If it is settled before the click, the whole chain behind it falls away: return shipping, inspection, refurbishment, putting back into stock and the loss in value on worn goods.

An honest classification is part of this: not every return can be avoided, and a selection order is not misconduct but a reaction to missing information. What can be influenced is the part that lies in one's own house, namely complete measurements, maintained fit notes and an assistant that reads out both at the right moment instead of estimating. How the moment in the cart itself is used is shown by the article Recovering cart abandonment with chat; which questions follow after the order, and what needs a connection to the ERP or inventory system, is covered by the article Order status and returns in the shop. The matching set of functions for the moment of purchase is bundled by the cart assistant.

What can be measured and what cannot

Size questions can be counted comparatively cleanly, because they almost always contain the same terms: size, measurement, fit, number, tight, wide. From conversation analytics you can read how often they are asked, which share of them could be answered from the stored chart and for which articles the information was missing. If the questions cluster around one manufacturer, a chart is usually missing there rather than a capability of the assistant. This turns the analysis into a tool for catalogue maintenance and not only for the chat; how to use that in daily work is explored in the article Analysing and optimising conversations.

• Share of conversations in which size or fit comes up at all
• Share of those that could be answered from the stored chart
• Articles and manufacturers without measurements, as a work list for catalogue maintenance
• Cases in which a comparison with an item already bought was possible
• Handovers to a member of staff, broken down by reason rather than only by count

What cannot be measured is the return that did not happen, because there is no control group for it. A particular reduction of the return rate therefore cannot be promised; it depends on range, target group, data quality and season. What remains is a factual chain: more complete measurements, an answer at the decisive moment, a named gap instead of an estimated figure, and a handover to a human as soon as a case falls outside the pattern. How that handover works cleanly is described in the article AI assistant: handover to human agents; how an assistant works in a shop overall is shown on the page assistant for online shops.

Frequent questions