- Baymard puts delivery being too slow at 20% of US shoppers who abandoned a checkout, against 40% who left over extra costs. Speed matters, price matters more.
- A late promise loses the sale. An early promise loses the customer. Those two errors are not symmetrical, and the models built for this are trained to treat them differently.
- If you state no shipping time in the United States, the FTC rule assumes 30 days and gives the buyer a cancellation right the moment you miss it.
- In the United Kingdom the default is also 30 days, and a consumer can end the contract and demand every payment back without undue delay.
- The honest estimate is the one with the buffer written into it before the cart, not the apology email sent after.
Every shop makes a promise it has only partial control over. The box leaves your hands on a Tuesday and arrives when a van, a depot, a weather system and a stranger's doorbell decide it arrives. In between sits a number on your product page that the customer reads as a commitment and you wrote as a guess.
The gap between those two readings is where refunds, support tickets and one star reviews come from. It is also, unusually for a logistics problem, something a very small seller can improve in an afternoon, because most of the error is not in the carrier network. It is in the parts of the estimate you own and have never measured.
What is an estimated delivery date actually made of?
Five components, only two of which belong to the carrier. The other three sit inside your own operation, which is why two shops using the same courier can quote the same lane and miss by two days in opposite directions.
Start with the cut off. An order placed at 16:55 against a 17:00 cut off behaves completely differently from one placed at 17:05, and most storefronts display the same estimate to both. Then pick and pack, which is the step small sellers chronically underestimate because they remember their fastest day rather than their median one. Carrier transit is the published service level, usually reliable in the middle of the distribution and unreliable at the tails. The final mile attempt adds a day whenever nobody is home. The buffer is the only honest lever you hold, and it is the one most shops set to zero.
Measure the first two for a month and you will usually find the fix. If your median dispatch is 1.4 days and your product page says same day, the carrier is not your problem. Our walk through of the small decisions between an order and a doorstep covers the operational half of this in more detail.
Why does a late promise cost less than an early one?
Because one costs you a sale you never had, and the other costs you a customer you already won. Researchers building these systems treat the two errors as different problems rather than as one error with a sign.
A 2021 paper from a fashion marketplace, Online Fashion Commerce: Modelling Customer Promise Date, sets this out directly. Promise a date too far out and the shopper may never place the order. Promise one that arrives before the parcel does and you get a poor experience the authors link to churn. Their response was to abandon ordinary symmetric error measurement and use loss functions that penalise the two directions unequally, comparing statistical methods against conventional machine learning and deep learning before deploying the result in production against an older rule based system.
The paper reports no accuracy figures, and it is worth being plain about that rather than inventing one. What it does establish is the design principle, and the principle transfers to a shop with no machine learning at all. Decide which error you can afford, then set your displayed date so that the error you can afford is the one you usually make.
For most small sellers that means quoting the date you hit roughly nine times in ten rather than the date you hit on a good week. The cost of that choice is visible: a slightly slower promise at the cart. The cost of the opposite choice is invisible until it arrives as a support ticket, a chargeback or a customer who simply never returns.
How late can you be before the law gets involved?
Sooner than most sellers assume, and the trigger is usually silence rather than lateness. Both major regimes default to 30 days and both hand the buyer a cancellation right that does not depend on your goodwill.
In the United States, the FTC guide to the Mail, Internet, or Telephone Order Merchandise Rule is explicit. State no shipping time and you must have a reasonable basis for believing you can ship within 30 days of a properly completed order. Miss the date you stated, or that default, and you owe the buyer a delay notice with a revised date and a free way to cancel. Where the new date slips by 30 days or less, silence can count as consent. Where it slips further, or where you cannot give a date at all, the order cancels automatically unless the buyer agrees. Refunds on cancellation run to seven working days for ordinary payment methods, and a credit toward a future purchase is not a refund.
The United Kingdom lands in a similar place by a different route. Section 28 of the Consumer Rights Act 2015 treats a contract with no agreed time as requiring delivery without undue delay and in any event within 30 days. If timing was essential, or if the consumer sets a reasonable new deadline and you miss that too, they may treat the contract as at an end, and you must reimburse everything paid without undue delay.
| Situation | United States, FTC rule | United Kingdom, s.28 CRA 2015 | What it means for your page |
|---|---|---|---|
| No shipping time stated | 30 days, needs a reasonable basis | Without undue delay, 30 days maximum | Silence is not safer than a date |
| You miss the date | Delay notice with revised date plus free cancellation | Consumer may set a new reasonable deadline | Build the notice before you need it |
| Delay runs long | Over 30 days late cancels automatically without consent | Contract may be treated as at an end | Long silences end orders by themselves |
| Money back | Seven working days, no vouchers | Reimburse all payments without undue delay | Store credit does not discharge it |
Neither regime asks whether the carrier let you down. The promise was yours.
Does a tighter date actually sell more?
Less than the pricing page does. In Baymard's ranked list of abandonment reasons, measured among US online shoppers once the just browsing group is excluded, delivery being too slow sits at 20%, while extra costs such as shipping, tax and fees lead the list at 40%.
Read those two numbers together before you spend money on speed. A free or flat shipping threshold moves twice as many people as a faster van. The common mistake is to buy a premium service to shave a day off the estimate while leaving a surprise fee at the final step, which trades the cheaper problem for the dearer one.
Speed still earns its keep in two places. It earns it on gifts and occasions, where the date is the product and a miss is total. It earns it on repeat categories where the customer is comparing you against a marketplace they already trust. Outside those, an accurate date beats a fast one, and the cheapest conversion work remains the part of the cart that mentions money. Our note on what cart recovery can and cannot fix makes the same argument from the other end of the funnel.
What should a small seller measure first?
Three numbers, none of which require software you do not already have. Your median hours from order to dispatch, your rate of missed first delivery attempts, and the share of orders that beat the date you displayed.
That third number is the one nobody tracks and the only one that scores the promise rather than the process. Export a month of orders, put the displayed estimate next to the actual delivery scan, and count. Anything above 90% means your estimate is honest and probably conservative. Between 70% and 90% means you are gambling with a minority of customers. Below 70% means the number on your product page is decoration.
Work the third number properly and it will surprise you. Take 200 orders, count how many arrived on or before the date shown at checkout, and express it as a percentage. A shop running at 78% is failing roughly one customer in five on a promise they paid against, which at a 60 pound average order and a 4% dispute rate is real money before you count the support time. Moving to 93% usually takes one change, not five: either a single extra buffer day on one slow lane, or an earlier cut off that stops marginal orders being counted as same day.
Address quality is the quiet multiplier underneath all three. A parcel sent to an ambiguous address fails its first attempt more often, and the failure lands in your support queue rather than the courier's. We made that case in detail when looking at why address data beats clever routing, and it applies just as hard to the estimate as it does to the route.
Carrier service levels describe the middle of a distribution, not a guarantee for your parcel. A two day service that hits two days 85% of the time is not a two day promise, it is a two day promise with a 15% failure rate attached. Quote from your own measured outcomes rather than from the courier's marketing page.
How do you write the estimate on the page?
As a range with a cut off attached, not as a single heroic number. A range tells the truth about variance, and the cut off converts a vague promise into a decision the shopper can act on right now.
Four rules cover nearly all of it. Show a window rather than a point, because every honest estimate is a distribution. Name the cut off next to it so the customer knows which side of the line their click falls on. Recalculate for the destination rather than quoting one national figure, since the lane to a rural postcode is not the lane to a city centre. Hold the same date through the cart and the confirmation email, because an estimate that moves after payment reads as a bait.
Exclusions deserve one line, not a paragraph. Heavy items, made to order products and remote destinations are the three that genuinely behave differently, and saying so plainly costs a sentence and prevents a dispute. If your parcel pricing is already pushing you toward awkward compromises, the mechanics in how couriers bill you for air explain why the heavy and bulky exceptions exist at all.
Why are cross border estimates so much worse?
Because a customs step is a queue with no published service level. Domestic transit varies by hours around a known median, while a border adds a wait that depends on paperwork quality, random inspection and a working day calendar you do not control.
The practical effect is that the same courier, same weight, same service tier will land inside a two day window at home and anywhere inside a two week window abroad. Shops that quote international delivery as a single number are usually quoting the lucky case, and the complaints that follow are not about the extra days, they are about the surprise. A window labelled as a window generates almost no tickets. A point estimate missed by nine days generates one every time.
Two things reduce the variance rather than merely describing it. Accurate commodity codes and declared values move parcels through faster because they do not invite a manual check, a point we went through in what the customs code on your parcel commits you to. Prepaid duties remove the second queue, the one where a parcel sits waiting for a customer to pay a fee they did not expect, which is the single largest contributor to the tail in consumer shipments. Neither is free. Both are cheaper than the refunds.
For a seller shipping a handful of international orders a month, the honest answer is often to quote generously and ship faster than promised. Delight costs nothing when it is the direction of your error. The same logic stretched over months is what a pre-order promise commits you to.
What changes when the dates are wrong at scale?
The support load arrives before the refunds do. Where is my order tickets are the single most predictable consequence of an optimistic estimate, and they peak two days after the date you printed, not on the date itself.
That lag is useful. It means a bad promise is measurable before it becomes expensive, if anyone is counting. Tag those tickets for one month, plot them against the displayed estimate, and the pattern usually points at one lane, one product class or one weekday. The fix is then specific rather than a blanket extra day added to everything, which is what shops reach for when they have no data and which quietly costs them the orders Baymard counted.
There is a second pattern worth tagging while you are in there. A meaningful share of those tickets are not late parcels at all, they are parcels delivered to a safe place the customer did not look in, or a first attempt made while nobody was home. Those arrive with the same wording as a genuine delay and they need a different answer, so counting them separately stops you adding a buffer day to fix a problem a photo on delivery would have solved for nothing.
Peak season deserves its own treatment. Carrier transit degrades predictably in the weeks before a major holiday, and the cut off you publish for that period should be derived from last year's actual scans rather than from the courier's published guidance, which is written for its network as a whole rather than for your lane. If you kept no scans last year, start keeping them now, because the estimate you can defend next December is the one you measured this October.
None of this needs a model. A spreadsheet of displayed dates against actual deliveries, kept for a month, will tell you more than any prediction tool you can buy, and it is also exactly the dataset such a tool would need before it could help you. If you want the storefront itself to show a destination aware window rather than a single flat figure, that is a build rather than a plug in, and it is the kind of change a store you can edit yourself absorbs in an afternoon.
The promise is the product, in the narrow sense that it is the only part of delivery the customer experiences before they pay. Make it slightly duller than you would like, keep it, and the rest of the logistics stack gets quieter on its own.