Why return reasons only help once they are coded
The returns policy and experience guide asks for the reasons to be coded and not only collected, as a short list with a free-text field, read by product, variant, campaign and channel. Other pages lean on the result. The diagnosis of a rising return rate cuts a rise by coded reason alongside product, variant, campaign and gift orders, and the first-product tutorial points a product with a low second-order rate and scattered gaps to its reviews and return reasons. Both assume the codes already exist.
None of them says how to get there: who should own each reason, how the list is drafted, where it is captured, what it is joined to, what it is counted against and how a cluster reaches the person who can change something. Without those steps, coded reasons are a column in a spreadsheet that leads to no work. What the finished measure answers is set out in the table.
| Question | What the table reads | Where the answer is used |
|---|---|---|
| Why do items come back? | One main reason for each returned item, from a short defined list | Choosing which fix to make first |
| Where do the reasons sit? | Each reason as a share of the units sold of the product, variant, campaign or channel concerned | The product page, the photography brief, campaign copy and the parcel |
| Who acts on it? | The owner named for each reason before the list was written | The brief each owner receives |
| Did the fix work? | The same cuts read again on a closed cohort, with the other reasons beside them | Keeping, changing or dropping the fix |
The tutorial stays inside the store's own records. It gives no legal advice either, so take advice on the rules of each market it sells in.
Before you start
- The returns export, from the returns or order system: the return date, the order number, the item returned, the quantity, any reason now held and any free-text note, for a recent run of returns.
- The order records: order number, order date, line, product, variant, quantity, the campaign and channel where the order system records them, the gift option chosen at checkout and a customer number.
- The units sold for the same products and variants over the periods to be read, because each count needs a base.
- The person who runs returns, who knows what the warehouse sees and what the returner is asked, and the owners of the pages, photographs, campaigns, product records and parcels, who will each receive a brief.
- A spreadsheet or SQL. Joining two lists and counting is enough.
- Time to read real returns. The list and the free-text reading both depend on a person reading actual items, not on a tool that summarises them.
The steps
Decide who acts on each reason before writing the list
As general practice (the site's guides do not state this order of work for returns), start from the people who can change something and work back to the reasons. Write down, for this store, who owns each fix the returns policy guide pairs with a cluster: whoever writes the product pages, whoever briefs photography, whoever writes campaign copy, whoever packs the parcel and whoever owns the returns policy. A reason with no owner is a count nobody will act on, so either find the owner or fold that reason into a catch-all reason with a note. Name one person who owns the list itself, its definitions and its changes, and who will sit with the returns lead when the counts are read. Keep the owners beside the draft of the list; the next step will adjust both.
Draft the list from real returns, with one main reason per returned item
Pull a sample of recent returns, spread across products, variants and weeks, with any reason and note already held. The tutorial on tagging support contacts gives the method for support: draft from real contacts, keep one main reason for each, give each reason a definition and an example, keep the reason apart from the root cause, and give the list an owner and a change log. The glossary entry for a codeframeThe defined set of codes open-text answers are tagged with, built from a sample, applied consistently, checked by a second coder and counted against a base. adds a second coder's check and a stated base for the counts, and a contact reasonThe main reason a customer contacted support, tagged by the agent from a fixed, defined list so volume and other service metrics can be read by cause. is the same idea applied to support. Follow them as written; what changes for returns is the unit and the owners. As general practice, record one main reason for each returned item and not for each order: an order can hold a jacket that was too small and a belt with a fault, and a single reason for the order would force a choice between them. Keep the list short enough to choose from at the moment of returning, and use the example list in the returns policy guide as a prompt and not as a template. The reason is what the returner says; the missing fit note, or a cut that runs small, is a cause found later, once the clusters are tested.
Capture the reason where the return is made, with a free-text field and the facts already known
Ask for the reason at the point where the return starts, on a page linked from the order email and from the account, as the returns policy guide describes, and, as general practice, ask once for each item being returned. The returner chooses from the short list and a free-text field takes the rest. As general practice, keep the free text optional and below the list, so that it adds detail and does not replace the choice, and do not ask the returner for what the store already holds. The order number, the product, the variant and the order date come from the order record, and so does whether the order carried the gift option. That option is the store's main record of a gift, as the guide to selling gifts online puts it, and gift returns are read apart because the giver's guess about someone else's taste drives them, not the product page. Also as general practice, show the list in the customer's words and not the store's, show the item beside it so the reason is chosen for that item and not for the parcel, and keep the same list wherever a return can start so that the counts add up.
Record what the warehouse finds on receipt apart from what the customer says
This step is general practice; the site's guides do not state it. When the parcel is opened, the warehouse sees what the customer cannot describe: the state of the item, whether the packaging held and whether the item is the one that was ordered. Record that in a field of its own, from a short list of conditions the warehouse can apply without judging the customer, with a note for anything that does not fit. Keep it apart from the reason the customer chose. The two answer different questions: the reason says what the customer believes happened, and the warehouse record says what arrived. Do not overwrite one with the other. Where they disagree, such as a parcel reported as damaged whose item shows no sign of it, the disagreement is something to read in the free text and the warehouse note, not an error to correct. The warehouse record is a check on the reasons it can observe, such as damage, and it does not decide what the store accepts back, which stays with the returns policy.
Join each return to the order line it reverses and keep one row per returned item
Settle the grainWhat one row of a table represents, such as one order, one order line, one session or one customer per day; queries and metric definitions must match it. of the table before building it: one row for each returned item, carrying the quantity returned so that units can be added up later. Then join each row to the order line it reverses. A join keyThe shared field, such as date and channel, a campaign ID or a click ID, on which rows from two data sources are matched in a blend or a query. has to mean the same thing, have the same type and sit at the same grain on both sides. As general practice, use the order number with a line identifier from the order system. The ecommerce tracking tutorial has the store send the order number as the transaction identifier and records, from Google's developer guide, how refunds reach analytics: with the original transaction identifier, and with only the refunded items when the refund is partial. Reduce each side to one row for each key before joining, as the SQL for marketers guide advises, so that a line of several units or a partial return does not repeat rows. Then check the join: the returned items it produces should add up to the returns system's own count, and each return that matched no order line is listed with a reason beside it, not dropped. The join brings in what a return lacks: product, variant, order date, campaign and channel where the order system records them, and the gift option, so that a return can be attributed to a channel, campaign or product, as the return rate entry says. Keep the analysis file to a customer number, and no names or addresses. The personal data in analytics guide lists an internal customer number among the identifiers that are still personal data under the GDPR, and data minimisation is the reason to leave out what the analysis does not need. This is general direction and not legal advice.
Count each reason against what was sold, once the return window has closed
A return rateShare of orders or units sold that customers sent back. can count orders or units, and the basis has to be stated beside the figure. Units suit a reason recorded for each item, because a basket can hold several items. As general practice, read each reason as a share of the units sold of the product or variant concerned: the units returned for that reason, divided by the units sold and shown per hundred units sold. Count against the orders placed in a period and not against the returns that arrived in it. Returns belong to the sale they reverse and arrive after it, so only a cohortA group of customers who share a starting event and period, tracked together over time. of orders whose return window has closed gives a final figure, as the return rate entry and the rising return rate diagnosis both say. Mark any cohort whose window is still open as provisional and keep it out of comparisons. Cut the table by product, variant, campaign and channel, as the returns policy guide asks, and read gift orders apart, as the diagnosis does. Show the units sold beside each rate. A variant that sold few units gives a rate that moves a long way with a single return, so combine small groups into one row or lengthen the period, as the reading small numbers guide recommends.
Read a sample of the free text and let a second person check the codes
Short codes hide detail, a warning the creative angles guide makes about return reasons in particular, so read the free text as well. Read a sample of the notes from the largest clusters and from any catch-all reason, and write down what they say in the customer's words: a note under "changed mind" that names a feature belongs to a different reason. The voice of the customer guide lays out how open text is read and how the loop is closed, and here the loop is the customer who explained the problem seeing it fixed. A second person codes the same sample and the two sets are compared, as the tagging tutorial's audit and the codeframe entry's second coder describe; only the material is new here. Where the notes show one theme under several codes, sharpen a definition or add a reason. Because the notes are text typed by customers, they can hold names, addresses and order references, and the personal data guide lists names, postal addresses and free text a person typed among what to keep out of analytics. Read them in the returns system or from a copy with identifying details removed, and quote the meaning and not the identity.
Route each cluster to its owner and test the stated reason against what customers did
Take the clusters the table shows: a reason whose rate for one product, variant, campaign or channel stands well above the rest of the range or above the store's own earlier closed cohorts. Send each to the owner named in the first step, with what the reason points to. A fit reason goes to whoever writes the product page, where the product page optimisation guide describes a size guide with measurements and guidance on fit, and puts fit feedback beside the size selector. A reason about the look of the item goes to the next photography brief. A reason concentrated in one campaign's orders goes to whoever wrote the campaign, who can set the ad's claim beside the page and the product, as the campaign merchandising guide asks the page to continue the ad's message. A damage reason, read with the warehouse notes, goes to whoever packs, and the packaging and order inserts guide puts damage first among what to read in the store's own returns data. As general practice, test the largest cluster in two ways before building the fix. First, run the five whys: state exactly what is happening, ask why, answer each why from evidence and stop at something the store can change, a page, a process or a check, and not at a person. Second, set the reason against what customers did next. The stated and revealed preference framework separates what people say from what their choices show, and a returner's chosen reason is a stated answer. An exchange, an order for another size of the same product, an order for something else or no order is behaviour. A too-small return followed by an order for the larger size supports the reason; one followed by nothing leaves it as a stated answer only. Where the two disagree, let the behaviour set the next question and read the free text again.
Write a brief for each owner and recheck after the fix
As general practice, write one short brief for each owner and not one report for the whole team. Each brief holds the cluster (the reason, the product or variant, the rate with the units sold beside it and the closed cohort it comes from), a few anonymised lines from the free text in the customers' words, the cause the testing supports, the change asked for and the date the cohort will be read again. The returns export is already an input to the photography brief tutorial, so for a cluster about colour or look the work is to hand over the cluster with its counts and not a general complaint. After the change, read the same cuts again on the next closed cohort. The rising return rate diagnosis gives the check: the reason that was fixed falls for those variants, and no other reason grows in its place. Mark the date of the change on the table so cohorts that straddle it can be told apart.
Review the list when it stops fitting the returns and keep a change log
The list will stop fitting at some point: a new product line brings returns the codes cannot describe, two reasons overlap, a catch-all reason grows or the second reader keeps disagreeing with the first. As general practice, review the list on a fixed cycle, as the tagging tutorial asks for changes, looking for those signs, and make any change as one dated revision, with the old reasons mapped to the new so earlier cohorts can be restated. Keep a change logA dated record of changes made to tracking, accounts, reports, lists or rules, with who made each and why, so later movements in the figures can be traced. with the date, the change, the reason and the owner, and keep it beside the list, as the change log entry describes for reason lists, adding an entry to the measurement log in the measurement change log tutorial when a change moves a reported rate.
Reading the reasons
Compare each row with the store's own earlier closed cohorts and not with another store's figures. The example below uses invented figures, so they set no target. Treat each reading as a hypothesis, which the eighth step tests for the largest cluster. The patterns for fit, look, campaign and gift orders are those the rising return rate diagnosis and the steps above already give, so the table below keeps to the ones they leave out.
| What the table shows | What it may mean | Owner | First thing to try |
|---|---|---|---|
| A damage reason concentrated in one product type, with matching warehouse notes | The parcel does not protect that product | Whoever packs | Read the warehouse notes and the packaging for that product |
| A large share under a catch-all reason, with free text that names a feature | The short list hides a specific complaint | The owner of the list | Read the notes, then add a reason or sharpen a definition |
| One reason falls after a fix while another rises | The fix moved the returns and did not remove them | The owner of the list and of the fix | Read the free text behind the rising reason and revisit the fix |
| A rate that rests on a handful of units | The count is too small to read | Whoever reads the table | Combine groups or lengthen the period before acting |
A knitted hat seller reads its returns (illustrative)
Yolvaxen Hats codes its first autumn of returns
Yolvaxen Hats, an invented online seller of hand-knitted beanies and scarves, has a returns form with one free-text note that nobody has coded. The returns lead and the product page owner read a sample of 90 past returns and draft six codes, each with an owner, as the table shows. The form now asks for one main reason for each returned item, the warehouse records condition on receipt in a column of its own, and returns are joined to order lines on the order number and line number, with a customer number and no names.
For an autumn cohort of 4,000 units whose return window has closed, 500 units came back. The table counts each reason per hundred units sold.
| Code | Owner | Units returned | Per hundred units sold |
|---|---|---|---|
| Fit too tight | Product page owner | 160 | 4.0 |
| Fit too loose | Product page owner | 40 | 1.0 |
| Colour differs from the photographs | Photography brief | 120 | 3.0 |
| Knit or finish fault | Production lead | 24 | 0.6 |
| Damaged in the post | Packing lead | 16 | 0.4 |
| Changed mind | Returns lead | 140 | 3.5 |
Cut by variant, 112 of the 160 fit-too-tight returns sit on the smaller size of the ribbed beanie, which sold 800 units: 14.0 per hundred against 1.5 for the rest of the range, and its page has no fit note. Of the 120 colour returns, 84 sit on two colours, moss and rust, which sold 700 units: 12.0 per hundred against 1.1 for the other colours. A second person codes a sample of 80 free-text notes and agrees with the first on 70; the ten disagreements sit between changed mind and fit too tight, so the fit definition gains an example. Setting the fit reason against what customers did next, 52 of the 112 ordered the same beanie in the larger size, which supports the reason for those 52 and leaves the other 60 as stated answers only, and a five whys on the cluster ends at a page template with no field for fit.
The store sends the first cluster to the product page owner, who adds a fit note beside the size selector, and the second into the next photography brief. When the next autumn cohort has closed, fit too tight on the smaller size falls to 6.0 per hundred, 45 returns on 750 units sold, while fit too loose and changed mind stay at 1.0 and 3.5 across the range and no other reason on that size has risen, so the store reads the fall as a likely fix and not as returns moving to another reason. The colour reasons wait for the reshoot and the cohort after it, and the change goes into the log with its date.
Other retention work, from flows to loyalty, sits together in the retention section.
Common mistakes
- Collecting reasons without an owner. A reason nobody can act on is a count that sits in a spreadsheet.
- One reason for a whole order. An order can hold items that came back for different reasons.
- Letting the customer's reason and the warehouse's note blur into one field. The disagreement between them is information.
- Counting reasons against returns and not against units sold. A reason can fill a large share of a product's returns while staying small against its sales.
- Reading a cohort before its return window has closed. Recent orders look better than they will turn out.
- Dropping returns that match no order line. They are a list of keys that failed to meet, not noise.
- Trusting the short code alone. The free text holds the detail the code hides.
- Putting names and addresses into the analysis file. Names and addresses add nothing the analysis needs, and a customer number is still personal data.
- Fixing the largest reason without testing it. A cluster can have a different cause from the one it first suggests.
- Changing the list without a note. Cohorts before and after the change stop being comparable.
