Returns: The Cheapest Fix Sits on the Product Page

Nineteen point three percent. That is the share of online sales retailers expected to come back in 2025, according to the National Retail Federation's returns figures published on 15 October 2025. Across all of retail the rate was 15.8 percent, worth $849.9 billion. The gap between those two numbers is the tax on selling to somebody who cannot touch the thing first.

A small seller reads that and reaches for the policy, because the policy is the lever nearest to hand. It is the wrong lever. Tightening a returns policy reduces returns by reducing sales, and it does it in the part of the funnel you paid most to fill. The same NRF work found 82 percent of shoppers treat free returns as a major purchase consideration and 71 percent are unlikely to come back after a bad returns experience.

The lever that works is upstream, and it is unglamorous: say more, earlier, about the thing you are selling.

Why is the product page the cheapest place to cut returns?

Because a return usually starts as a wrong expectation, and expectations are set on the page. By the time a parcel is travelling back to you, every cost has already been incurred.

Think about what a single return actually consumes. You paid to ship it out. You pay, or the customer pays and resents it, to ship it back. Somebody opens the parcel, inspects the item, decides whether it can be resold, relists it or writes it off, and processes a refund. The stock was unavailable to anyone else for the fortnight it spent in transit. None of that is recoverable by being strict at the end. It is only avoidable at the start.

The research on how shoppers judge products is blunt about where the gap is. Baymard's work on in scale product images found that 42 percent of users try to work out how big something is from the photographs, while 28 percent of large ecommerce sites provide no image showing the product in context at all. One test participant put it plainly: without context they could not tell how big it was. A shopper who guesses the size wrong either abandons a product that would have suited them, or buys it and sends it back.

Can you simply refuse more returns?

Not in most of the markets a small shop sells to, and not for the reasons that generate the most returns. The right to change your mind is written into law rather than into your policy.

In the European Union a distance sale carries a withdrawal right of 14 days without the buyer having to give a reason, counted from delivery for goods. The customer usually pays return postage, unless you offered to cover it or failed to tell them before the sale that they would pay. If the goods are defective, you cover the return cost.

There are genuine exceptions and they are worth knowing precisely, because several of them describe products small makers sell every day: customised or made to order goods, perishable items, sealed media once unsealed, and streamed digital content once consumption has started. If you make to order, the withdrawal right may not apply, and that is a material difference between your shop and a reseller's. Saying so clearly on the page is both a legal obligation and a way of setting expectations before somebody orders a personalised item on a whim.

The important implication for planning is simple. Your floor is not zero. Some share of returns is a feature of the market you chose. What you control is the share caused by your own listing.

Which return reasons are worth fixing first?

The ones that are both frequent and expensive, which is not the same as the ones that are loudest. Ranking needs two numbers per reason, and most shops hold neither.

This table is the working version, assembled from the cost structure of a small shop rather than from any single report. The point of the last column is to be honest about where automation earns its place.

Look at the fourth row for a second. Damage returns feel like a listing problem when the complaints arrive, and they almost never are. They cluster by packaging and by route, which is why the first useful analysis is a cross tabulation rather than a rewrite. We looked at the shipping side of this in our piece on how dimensional weight shapes packaging and shipping cost, and the same box decision often sits behind both the cost and the damage.

How do you capture the reason without annoying the customer?

With a short closed list at the moment of the refund request, four to six options, plus an optional free text box you do not analyse manually. The list is the data. The text box is the courtesy.

Most shops do the opposite. They offer an open field, collect three hundred sentences a year, read none of them, and conclude they have no data. A closed list gives you a countable variable from day one, and a year of it is worth more than any tool you could buy.

Two rules make the list useful. Keep the options mutually exclusive, because a shopper who could pick either will pick the first. And separate did not fit from not as described, since they point at entirely different fixes. If you sell apparel, split fit into too small and too large, which turns a vague complaint into a sizing correction you can actually make.

This is where a model genuinely helps, and it helps at the point most people skip. Once you have a year of closed list reasons plus free text, a model can read all of the text against the structured reason and tell you where the two disagree. Customers who picked changed my mind and then wrote three sentences about the colour are telling you something your chart is hiding. That analysis is a morning's work now and was a research project five years ago.

What the page has to say that it probably does not

Measurements, in the units your buyers use, per variant rather than per product. A size chart at the bottom of the page that covers your whole range is a chart nobody reads and everybody misreads. The measurement that prevents a return is the one attached to the exact item in the basket.

Scale, shown rather than stated. The Baymard finding is that people judge size from photographs whether or not you gave them a chart, so a photograph with a hand, a room or a familiar object in it does work no chart can. For a maker with a phone, that is one extra shot per product.

Material and finish in plain words, including the unflattering ones. Describing a fabric as substantial when it is stiff buys you a sale and costs you a return plus a review. The commercial arithmetic is not close: a return costs the shipping twice and a bad review costs indefinitely.

Delivery dates you can hit in a bad week rather than a good one. Late arrival returns are returns you caused by optimism at the checkout.

If you are building the shop rather than patching it, this is easier to get right at the start than to retrofit, because the fields have to exist before anything can fill them. An AI ecommerce store builder that generates the product schema alongside the pages gives you per variant measurement fields from day one instead of a description box you later have to parse.

Where AI actually earns its place

In reading, not in writing. The instinct is to point a model at your product copy and ask for better descriptions. That produces fluent text and no change in your return rate, because the missing information was never in the shop to begin with. A model cannot know the inside leg measurement of a trouser you never measured.

What it can do is find the pattern. Give it your order lines, your return reasons and your product attributes, and ask which attributes correlate with returns. Small catalogues surface obvious answers fast: one supplier, one colourway, one size that runs small across every style. Those are findings you act on in an afternoon.

It can also draft the text once you supply the facts. Handed a spec sheet with real measurements, a model writes the per variant fit note for two hundred products in the time it takes to make coffee, and your job shrinks to checking a sample. That is a good division of labour: you own the facts, the model owns the tedium. Our method for writing product descriptions with AI sets out the same boundary in more detail.

And it can handle the return once it happens, which is a different problem from preventing one. Deciding refund against exchange, routing an item to resale or repair, answering the where is my refund email: those are rules driven tasks with clear success criteria. We covered that side in what to automate in returns and refunds. Automating the handling lowers your cost per return. Only the page lowers the count.

Bracketing, and why your best customers return most

Bracketing is the habit of ordering two or three sizes with the intention of keeping one. It is rational behaviour for a shopper who cannot try things on, and it inflates a return rate without indicating anything wrong with the product.

The NRF figures give a hint of where it concentrates. Shoppers aged 18 to 30 averaged 7.7 online returns in the previous twelve months, more than any other age group. That is not a generation of difficult customers. It is the cohort most comfortable with distance buying, doing the thing distance buying invites.

The commercial mistake is to treat a bracketing customer as a bad customer. Someone who orders three and keeps one has a higher return count than someone who orders one and keeps it, and quite possibly a higher lifetime value. If you tag returns by customer and look at net spend rather than gross returns, the picture often inverts. Deciding to discourage bracketing is a legitimate choice, but make it with that number in front of you, not from the frustration of a Monday morning parcel pile.

The preventative fix is the same one as everywhere else in this piece. People bracket when they cannot tell which size to order. Per variant measurements plus an in scale photograph reduce the need to hedge, which is a better outcome than a restocking fee, since it costs the customer nothing and costs you one afternoon.

What about the parts you cannot photograph?

Weight, smell, sound, stiffness and warmth are common return causes and none of them appears in a picture. The honest answer is to write them down in numbers or comparisons rather than adjectives.

A candle described as having a strong scent tells a buyer nothing. A candle described as noticeable across a small room within twenty minutes tells them whether to buy it. A jumper described as warm competes with every other jumper. A jumper described as suitable down to about five degrees with a shirt underneath sets an expectation that either matches or does not.

These comparative sentences are the single most underused tool on a small shop's product page, and they are also where a model is useful in a way that does not risk invention. Give it your product attributes plus the comparison you want drawn, and it will turn a specification into a sentence a person can act on. What it must not do is supply the specification. A model asked how warm a jumper is will answer, confidently, about a jumper it has never touched.

One more category deserves a mention because it generates disputes rather than plain returns: anything where the buyer's expectation was set by a photograph on a screen they have not calibrated. Colour is the obvious one. Naming a colour in plain words next to the marketing name, and saying which screen the photograph was taken on, costs a line and prevents the argument that starts with the word nothing like.

How would you know it worked?

By comparing products you changed against products you did not, over the same period. Return rates move with season, with promotion and with product mix, so a before and after on the whole shop tells you almost nothing.

Pick the twenty products that generate the most return cost. Fix the pages for ten. Leave ten alone. Wait one full return window past a normal sales cycle, which for most shops means a quarter rather than a fortnight, then compare the two groups. If the fixed group did not move, your theory about the cause was wrong, which is worth knowing before you rewrite four hundred more listings.

Measure cost rather than count, too. A shop that cuts returns on a nine pound item while ignoring an eighty pound item has improved a statistic and not a business. Multiply each return by what it actually burns: the outbound postage you already spent, the inbound postage, the handling minutes, and the share of items that come back unsellable.

The honest summary

Online returns run several points above the retail average and that will not change, because the shopper cannot hold the product. Within that, a meaningful share is self inflicted: a chart that was generic, a photograph with no sense of scale, a delivery date set by hope. Those are fixable with a tape measure, a phone camera and an afternoon per twenty products.

The customer who returns something is not the enemy in this. They bought once and they were willing to buy again until the experience taught them otherwise. Every source on this agrees that the return experience predicts the next purchase. Keep it easy, and spend the effort making the first description true enough that fewer people need it.