Why the share of promises kept is the figure to build
The delivery promises guide treats the delivery promiseA store's claim about what delivery will cost and when an order will arrive, made on product pages, at checkout, in confirmations and in ads. as a claim that operations set and marketing states, made of two parts operations own: the store's time to prepare and send an order, and the delivery service's time from dispatch to the door. It then asks for the promise to be measured as the share of orders that arrive on or before the date the customer saw when ordering, split by option, product type and area. Its list of mistakes names the alternative that goes wrong here: tracking how fast parcels travel instead of whether the promise held. A parcel that takes a given number of days is on time under a longer promise and late under a shorter one, and the expectation-disconfirmation framework is the reason the customer's verdict follows the promise and not the clock: the experience is judged against the expectation the business set.
Several pages assume the measure exists. The guide's glossary entry and its own measuring section define it, the online store audit makes the promise a gate, and the peak season trading playbook lists the delivery promises kept among the outcomes a peak is judged on. None of them builds it, and building it takes decisions that are easy to get wrong without noticing, such as what to do with an order that never arrived. The table sets out what the finished measure answers.
| Question | What the measure reads | Where the answer is used |
|---|---|---|
| Is the promise kept? | The share of orders arriving on or before the promised date, with the orders that never arrived counted as missed | Whether to leave the promise alone, split it or lengthen it |
| Where does it fail? | The same share by dispatch or delivery leg, option, product type, area and week | Which part of operations fixes it: dispatch, the delivery service relationship or the promise itself |
| What follows from a miss? | Contacts about delivery per order shipped, abandonment around a change, repeat orders by delivery experience | How urgent the fix is against the store's other work |
| Is it staying fixed? | The share by week, read against its own history | The review with operations |
This page measures the promise and says little about how to word it. Where the promise is shown, and how it is kept the same across pages, feeds and emails, stays with the guide. This page gives no legal advice, so take advice on the rules of the markets the store sells to.
Before you start
- The order records, from the order system: order number, the time the order was placed, the delivery option the customer chose, the destination area, the product types in the order and its status. One order is one row.
- The promise as shown at the time. Ideally the promised date or window is saved on the order when the customer pays. If it is not, find the wording and the cut-off rule that were live on each order's date, which is a task for the 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. the store keeps or the page history; the step below turns them into a reconstructed date that is labelled as one.
- Dispatch records: when each order left the building. If the store cannot supply them, the total can still be measured, but not the split between the store and the service.
- Delivery records: the date each parcel was delivered, as the delivery service reports it. The post-purchase email flow tutorial notes that the delivery event is the one most shops lack and asks whether the shop or the carrier integration can send it, which is worth finding out now because it feeds both this table and the delivery message.
- A period long enough to be ordinary. Choose the window before looking at results, include ordinary weeks and mark any peak or sale weeks so they can be read apart, and keep the window the same at each rerun.
- Someone in operations who owns the promise and will read the table with you, because each pattern in the table points to an owner.
- A spreadsheet or SQL. Joining three lists on an order number and counting is enough; the SQL for marketers guide shows how to avoid the repeated rows a join can create.
The steps
Write down the promises as customers saw them and decide what kept means
List each promise in force during the window, by delivery option, product type and destination, with the date each one changed. Then write one definition of "kept" and apply it to each order: delivered by the end of the promised day, or by the last day of a promised window; counted in working days or calendar days, in the way the checkout counted them; and "delivered" meaning the status the delivery service reports and not the customer's own receipt. Write it next to the owner's name and keep it. Changing the meaning of kept halfway through a series makes all earlier figures incomparable, so the metric definition change checklist applies if it ever has to change. Items that are made to order carry a promise of their own, with the making time stated apart from delivery, as the guide to made-to-order and personalised products describes; list their promise separately.
Export one row per order with the promised date saved at the time of ordering
Take the order records and add, for each order, the date or window the customer was shown. The guide and its glossary entry ask for the promised date to be saved with the order when it is placed, not worked out again afterwards, because a date adjusted after a delay conceals the miss. Where the store did not save it, reconstruct it from the wording and cut-off rule that were live when the order was placed, working the date out the way the checkout would have, and mark each such row as reconstructed. This is general practice, not a substitute: start saving the promised date on each order now, so that the next run does not need the shortcut. Write down 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, one row per order, because the next step brings in records that can have several rows for one order.
Add the dispatch and delivery dates and check the join before trusting it
Match the dispatch records and the delivery service's reports to the orders on a field they share, such as the order number, which is the 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.. A split order can have several parcels and so several delivery rows; reduce each source to one row per order first, and as general practice keep the latest delivery date and a count of parcels, so the join does not repeat orders and inflate the counts. Then reconcile: the number of orders after the join should equal the order system's number for the window, and the orders with no dispatch date or no delivery date should be listed, not dropped. Look for reasons such as orders cancelled before dispatch, order numbers entered differently and parcels that were lost. As general practice, do not move on until each order without a delivery date has a reason beside it.
Remove what has its own promise or never shipped, and decide how split and mixed orders count
As general practice, take out the orders that never shipped because they were cancelled or refunded before dispatch and before their promised date had passed, since no promise was tested; an order that never shipped and was not cancelled stays in and is counted as missed once its promised date has passed. Measure made-to-order items in a table of their own against their own dates. Decide once how an order whose items carry different promises counts. The guide's example shows both promises on such an order. This is general practice, not something the site's guides state: count the order as kept only when each part arrived by its own promised date, keeping each part's promised date and delivery date in a small table of its own for these orders, and record how many orders fall under the rule. An order placed after the cut-off was shown the later date, so judge it against what it was shown and do not adjust it. Write each rule into the definition from the first step.
Mark each order kept, missed or still open, and count the orders that never arrived as missed
An order delivered on or before its promised date is kept. An order delivered after it is missed, and the number of days late is worth recording. As general practice, an order whose promised date has passed with no delivery record, unless the check in the third step finds its delivery date, is missed as well: a table that drops rows with an empty date flatters itself. Likewise as general practice, an order with no delivery record whose promised date has not yet arrived is open, and stays out of the rate until it does; the guide to comparing periods advises marking recent days as provisional because the latest days are incomplete. Report the open count beside the share so that a recent period is not read as better than it will turn out.
Split each miss into the store's part and the delivery service's part
This step is general practice, not something the site's guides state, though it follows the guide's two-part promise. For each order compute two intervals: from payment to dispatch and, where there is a delivery date, from dispatch to delivery. Compare the first with the dispatch time and cut-off the promise assumed, and the second with the delivery time the promise assumed for that service and area. A missed order is then one of three kinds: late leaving the building, on time leaving and late arriving, or both, and in the table an order late at both stages is counted under late leaving. An order with no delivery date has no second interval: if it left on time it counts as late arriving because it never arrived, and one that never left counts under late leaving. The kinds point to different fixes: dispatch is operations inside the building, delivery time is operations' relationship with each service, and a promise whose assumption was too tight for either is the promise's own fault. The split sits beside the headline figure and does not replace it, because the customer was promised a date by the store.
Read the share by option, product type, area and week, with the counts beside it
A store-wide share hides where the misses happen, as the guide warns, so cut it three ways the guide names: by the option the customer chose, by product type and by where the parcel went. As general practice, not something the site's guides state, cut it by order week too, and by the hour of the day an order was placed where the cut-off matters. Show the number of orders next to each share and combine small groups, which the guide to reading small numbers recommends, because a share from a handful of orders moves a long way with each one. Also as general practice, read how late the misses were as well as how many there were: a promise missed by a few hours is a different problem from one missed by days. Keep peak and sale weeks in their own rows, since blending them with ordinary weeks makes both unreadable.
Add the customer's side: contacts, abandonment and repeat orders
The guide names contacts about delivery and repeat purchase by delivery experience as readings beside the share kept, and asks for conversion and abandonment to be watched when a promise changes; the table can carry all of them. Count support contacts about delivery per order shipped, which is a contact rateSupport contacts in a period divided by orders, active customers or bookings in the same period, showing how often customers need help. read by 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.; the tutorial on tagging support contacts shows how to build the list that makes the count possible, and the customer service metrics guide explains reading it. Add the conversion rate and the cart abandonment rate for the weeks around a promise change, and the repeat purchase rateShare of customers who placed more than one order in the period. among first-time customers, split by whether the first parcel was on time or late, over the same fixed window for each customer and counting only customers whose window has closed, as the repeat purchase rate entry describes, with the caution the guide gives. The list of customers a delay affected, which the proactive service messages guide tells the store to keep, also shows how many of them still had to get in touch.
Choose the change by the pattern, and make it everywhere at once
Match what the table shows to an owner and a change, using the reading table below. Where the answer is a new promise, set it at the point where nearly all orders are met rather than at the typical day: read the share of orders delivered within each number of days for that option and area and pick the interval nearly all of them meet. If dispatch is falling behind, the guide's rule is to lengthen the promise for new orders while the queue is still short, and the guide to marketing within capacity lists shifting, slowing, queueing and stopping demand as the other choices. Then change the promise in each place it is made on the same day: the pages and checkout, the shipping settings in the feed (the product feeds guide lists shipping as an attribute in Google's product data specification), the shipping details in structured data (see the product structured data tutorial), the ads, the scheduled emails and the wording support uses. Tell the customers a delay affects before they ask, using a service message that carries no promotion, and log the change.
Keep the measure running, and keep its definition fixed
As general practice, chart the share kept by week for each delivery option and read it against the store's own history, with the orders behind each week beside it and options with few orders combined, as the small numbers guidance above says. The control charts framework turns a weekly figure into a line with limits taken from its own ordinary weeks, so one slow week is not read as a trend and a real shift is not explained away, and it asks for seasonal weeks to be handled apart. Log each promise change and show it beside the chart, as the measurement change log tutorial describes for entries shown beside a time series. Also as general practice, pair the share kept with conversion and abandonment as a guardrailA metric an experiment must not worsen beyond an agreed bound, watched beside the primary metric that decides the test.: a promise lengthened to make the share look better can lose orders, and the figure alone would call that a success. Rerun the full table after any change to the promise, a delivery service, the warehouse process or the calendar, review it with the owner in operations at an interval the two of you agree and keep each run for comparison. Before a peak, take the last order dates from these records, as the selling gifts online guide asks, and not from hope.
Reading the result
Judge each row by the store's own earlier runs and not by another store's figures; the figures in the example are invented and set no target. The patterns below are examples of what the table can show. The conversion section places the promise beside the other things that decide whether a visitor orders.
| What the table shows | What it might mean | Owner | First thing to try |
|---|---|---|---|
| Misses concentrate in dispatch, on certain weekdays or in the hours before the cut-off | The cut-off or the dispatch time assumed by the promise is tighter than the operation | Operations | Move the cut-off or lengthen the dispatch part of the promise for those orders |
| Misses concentrate in one delivery service or one destination area | The delivery time assumed for that service or area is too short, or the service is under-performing | Operations with the service | Give that area or service a promise of its own, or change the service |
| Misses concentrate in one product type | Those items ship differently from the rest, from another warehouse, by another service or made to order | Operations, with whoever owns the product record | Give the product type its own promise, shown on its pages |
| Misses appear only in peak weeks | The promise was set on an ordinary day | Operations and the trading lead | Set last order dates from the dispatch and delivery records and lengthen the promise while the queue is short |
| Misses are spread evenly across everything | The promise was set at the average, so it fails on a large share of orders | The people who set the promise | Reset it where the store meets nearly every order |
| Contacts about delivery rise while the share kept holds | Customers cannot find the answer, or order messages do not carry the date | Marketing and support | Check the order, dispatch and delivery messages and whether customers can find the answer for themselves |
| The share is high but conversion or abandonment worsened after a longer promise | The fix traded orders for a number | The people who set the promise | Test a shorter promise for the groups that meet it |
Two cautions sit behind the table. A share is only comparable with another when both have had the full time to close, so a recent week with open orders cannot yet be judged against an older one. The table records what happened under the current promises, carriers and processes; the next run shows whether a change worked.
A picnic goods seller checks its promises (illustrative)
Drazukel Picnic Goods reads twelve ordinary weeks and three busy ones
Drazukel Picnic Goods, an invented online seller of picnic sets, cool bags and folding camp tables, promises standard delivery in two to three working days for orders placed before 2pm and an express service the next working day. Its checkout started saving the promised date on each order four weeks ago, so it reconstructs the earlier weeks from its page history and marks them. For twelve ordinary weeks it takes 6,400 orders, joins its dispatch records and the delivery services' reports on the order number after reducing each to one row per order, and lists 140 orders with no dispatch date (cancelled or refunded before dispatch and before their promised dates, taken out) and 41 with no delivery date whose promised date has long passed (counted as missed). That leaves 6,260 orders, none of them still open.
| Group | Orders | Kept | Late leaving the warehouse | Late arriving |
|---|---|---|---|---|
| Standard, mainland | 4,310 | 94% | 2% | 4% |
| Standard, remote areas | 560 | 62% | 2% | 36% |
| Express | 1,390 | 89% | 8% | 3% |
| All orders | 6,260 | 90% | 3% | 7% |
An order late at both stages is counted under late leaving. The store-wide 90% hides two different problems. The remote-area misses are almost all in delivery, with the service taking longer than the promise assumed, and the express misses are mostly in dispatch, concentrated in orders placed in the last hour before 2pm, which left after the courier's collection. Delivery contacts run at 14 per 100 orders in the remote areas against 5 on the mainland. The three busy weeks around a bank holiday, read apart, keep only 71% of 2,040 orders.
With operations the store moves the express cut-off to noon, reads the share of remote-area parcels delivered within each number of working days (nearly all arrive within five) and gives those areas a promise of four to five working days in the pages, checkout, feed shipping settings, markup, ads, order emails and support wording on the same day, logging the change. It sets last order dates for the next bank holiday from the dispatch records. Ten weeks later remote areas keep 94% of 470 orders and express 96% of 1,160, delivery contacts in the remote areas fall to 7 per 100 orders, and conversion in those areas is unchanged, which suggests the longer promise did not cost orders.
Common mistakes
- Tracking how fast parcels travel instead of whether the promise held. A customer compares the parcel with the date shown, not with a stopwatch.
- Recalculating the promise after a delay. A date recomputed later hides the miss; save the date shown at the order.
- Dropping orders that have no delivery date. A table that drops rows with an empty date flatters itself.
- Judging a split order on one parcel. Decide once how orders with several parcels or several promises count.
- Reading a single store-wide share. It hides the option, area, product type or hour where the misses sit.
- Blending peak weeks with ordinary ones. They are not the same process; read them apart.
- Reading the open orders as kept. An order with no delivery record whose promised date has not arrived yet has not been tested.
- Fixing the figure by lengthening all promises. Watch conversion and abandonment beside the share, or the number improves while the store loses orders.
- Changing what kept means without a note. Earlier and later figures stop being comparable.
- Acting on a handful of orders. Combine small groups, or hold the difference as a question until the next run.
