What a promotion has to be judged on
A promotion is an investment: the business gives away margin, and usually spends on media, in exchange for something it would not otherwise have had. The judgement is whether it got that something, and at what price. There are four candidates, and a promotion should be judged on the ones it was run for.
- Extra contribution over the whole period around the sale, after the discount, the media and the dip that follows. This is the test every sale has to face.
- New customers who would not otherwise have bought, and who go on to buy again.
- Clearing stock that would otherwise be written down, where the comparison is against the write-down, not against full-price sales.
- Reactivating lapsed customers, measured by their later purchases.
Deciding which of these a promotion is for belongs before it runs, and the guide to planning promotions and discounts sets the job, the mechanic, the eligibility and the break-even uplift that this analysis later checks against. Writing the baseline, the measures and the readout date into the measurement field of a promotion brief before launch fixes the method before anyone sees the result.
The analysis window matters as much as the measure. It starts before the sale, because customers who know a sale is coming wait for it, and it ends well after, because customers who bought in the sale do not need to buy again for a while. A report that covers only the sale days sees the peak and misses both troughs.
Collect only the tracking the analysis needs. Google's e-commerce developer guide documents a coupon field on the purchase event and on items, an item discount field, and promotion events that record when a promotional banner is viewed or selected; the e-commerce tracking tutorial shows them in an implementation. Google's URL builder documentation describes the campaign parameters to put on every promotional link, so that the email, social and paid traffic the sale generates arrives under one campaign name; the naming conventions guide sets the spelling. The order system remains the record of discounts, margins and customers, and the sale promotion readiness checklist makes sure these are in place before the sale starts.
Building the baseline
The baseline is an estimate of what would have happened without the promotion. Everything else in the analysis is the difference between reality and the baseline, so its quality decides the answer. Without a randomised control there is no single right baseline, only several reasonable ones, and the honest approach is to build more than one.
| Baseline | How to build it | Strength | Weakness |
|---|---|---|---|
| Same period last year | Last year's orders for the same days, adjusted by the trend between the two years | Captures SeasonalityPredictable, repeating changes in demand tied to the calendar, such as months, holidays, weekdays or hours. | Holidays and paydays move; last year may have had its own promotion |
| The weeks before | Daily orders in a clean run of weeks before the sale was announced | Reflects the current trend | Misses seasonal change between those weeks and the sale |
| Unpromoted products | Sales of products the sale did not cover, scaled to the promoted range | Same customers, same days, same weather | Distorted if unpromoted products are substitutes for, or bought with, promoted ones |
| Unpromoted regions or channels | Markets, stores or channels where the sale did not run | Closest to a control | Rarely comparable; customers cross over |
Three rules make any baseline better. Exclude the period after the sale was announced, since waiting customers depress it. Exclude other unusual events (a stock-outA period when a product is out of stock and cannot be sold, which caps sales and should stop advertising for that product until stock returns., a site outage, a competitor's sale) or adjust for them explicitly; the measurement change log is where such events should already be recorded. And write the method down before reading the result, so that the baseline is not chosen because it gives the answer people hoped for. For events that recur every year, align the same-period comparison to the event's own dates rather than to calendar weeks, the same rule the seasonal sale pages guide applies to search traffic for the event page. The SOP for ending a sale or promotion logs the exact start and end times and any early stop as part of the close-down, so the dates the baseline excludes are on record.
Incremental orders and revenue
Incremental orders are actual orders minus baseline orders, summed over the whole window: before, during and after. During the sale the difference is positive; before and after it is usually negative. Report the three parts separately and the net beside them, because the during figure alone is what the business will want to quote.
Revenue follows the same arithmetic, with one complication. Sale orders carry discounted prices and often a different AOVAverage revenue per order., because buyers add items to reach a threshold or trade up to products that are now affordable. Compute incremental revenue from actual revenue against baseline revenue at the prices the baseline orders would have paid, not by multiplying incremental orders by an average. Units deserve a separate line for businesses whose customers buy consumables, since a rise in units per order is the first sign of stock-up.
Treat the estimate as an estimate. This is not an IncrementalityConversions caused by marketing. test in the sense the incrementality testing guide uses, where a randomised holdout supplies the counterfactual; it is a comparison with a modelled baseline, and its error is the error in the baseline. When the actual orders fall short of the plan rather than the baseline, the diagnosis for a sale or promotion underperforming checks the target first and then the offer, the campaigns, the list, the stock and the demand spent before the window.
Pull-forward, stock-up and cannibalisation
Most of the gap between the headline and the real gain comes from three effects.
Pull-forward is demand borrowed from the future. Customers who would have bought in the next few weeks buy during the sale instead, and the weeks after the sale fall below the baseline. Measure it as the shortfall against the baseline in the period after the sale, over a window set from how often customers normally buy. The same effect runs backwards when a sale is expected: customers delay purchases they would have made, and the weeks before fall below the baseline.
Stock-up is the same effect within an order. Customers buy more units than usual, then take longer to come back. It shows in units per order for existing customers during the sale and in the time to their next order afterwards, compared with their usual interval.
Cannibalisation is the sale replacing full-price purchases. It happens between products (buyers choose the discounted pan instead of the full-price one beside it), between customers (people who would have paid full price use the code because it was there, including codes that spread to voucher sites), and between channels (a sale email takes credit for orders the customer was about to place). Read promoted against unpromoted product sales, the share of code redemptions from customers who were already mid-checkout, and redemptions of codes outside the channels they were issued in.
None of these is a sign that the sale failed. They are the part of the headline that was never extra.
Margin and the cost of the discount
The decisive figure is contribution: revenue after the discount, the cost of goods, delivery, payment fees and returns, before fixed costs. The contribution margin entry defines it, and it changes in a sale in a way revenue hides. The discount comes entirely out of margin, so a modest price cut removes a large share of the profit on each order, as the ROAS calculator page for discount promotions explains.
Split the discount into two parts. The discount given on orders that would have happened anyway (the baseline orders that took place during the sale at the lower price) is a pure cost: margin handed to customers who would have paid more. The discount on genuinely incremental orders is what bought them. Then build the net figure:
- Contribution earned on all orders in the sale period, at the prices actually paid.
- Minus the contribution the baseline orders would have earned at full price.
- Minus the contribution lost in the dips before and after.
- Minus the media spend and any extra costs the sale created, such as additional fulfilment or a rise in returns.
What remains is the promotion's net contribution. The break-even calculator helps check how many incremental orders a given discount needs before it pays; compare that number with what the sale actually delivered.
Who the promotion brought in
A sale that loses money on the week can still be worth it if it brought new customers who stay. That claim has to be measured, not assumed.
Start by splitting the sale's orders by customer status from the order system: first orders from new customers and orders from existing ones. The guide to new and returning customers explains why this comes from the business's own records rather than from analytics' new users. Compare each group with its baseline, because a sale that mostly brought existing customers forward has a very different value from one that recruited new ones.
Then follow the new customers. Compare the Repeat Purchase RateShare of customers who placed more than one order in the period. of customers whose first order was in the sale with that of customers who first bought at full price in a comparable period, at the same age. Look at what they bought next: if their second order came in the next sale, the promotion may have recruited customers who buy only at a discount. For email and loyalty promotions, a Holdout GroupA randomly chosen part of an audience kept from exposure to measure what exposure caused. of customers who were not sent the offer, set aside before the send, turns this part of the analysis into a proper comparison; the readiness checklist is the place to plan one.
A promotion scorecard
The scorecard puts every part of the judgement on one page, with the method beside each line so that a reader can see what the figures rest on.
| Line | How it is calculated | Source |
|---|---|---|
| Headline orders and revenue | Actual totals during the sale | Order system |
| Baseline | Each method's estimate, stated separately | Order history, unpromoted products or regions |
| Incremental orders, during | Actual minus baseline in the sale period | Order system |
| Dip before and after | Baseline minus actual in the windows either side | Order system |
| Net incremental orders | During, minus both dips | Calculated |
| Discount on baseline orders | Discount given on orders that would have happened anyway | Order system, baseline |
| Net contribution | Sale contribution, minus baseline contribution, minus dips, minus media and extra costs | Order system, finance, ad platforms |
| New customers | First orders in the sale, against the baseline for first orders | Order system or CRM |
| Later behaviour | Repeat purchase of sale-acquired customers against full-price-acquired customers at the same age | Order system |
| Uncertainty | The range of net contribution across the baselines | Calculated |
State the result as a range, with the baseline that gives each end, and name what would change the conclusion. "The sale added between a small loss and a modest gain in contribution, mostly from existing customers" is a finding a business can plan with; a single precise figure for the gain is a guess that looks like a fact.
Zestbarrow Kitchenware's spring sale
Zestbarrow Kitchenware is an invented online kitchenware shop. Its ten-day spring sale, at 10% off most of the range, doubled daily orders from a baseline of 120 to 240, and the first report called it the best week of the year. The analyst builds the baseline three ways (last year adjusted for trend, the four weeks before the announcement, unpromoted lines) and reads a window from a week before the sale to a fortnight after. Each full-price order earns £22 in contribution; each sale order earns £17 after the discount.
| Line | Central baseline (120 a day) |
|---|---|
| Sale orders, ten days | 2,400 against a baseline of 1,200 |
| Dip in the week before | 70 orders (110 a day) |
| Dip in the fortnight after | 350 orders (95 a day) |
| Net incremental orders | 780 |
| Contribution from sale orders | £40,800 against £26,400 baseline |
| Discount given on baseline orders (within the line above) | £6,000 |
| Media | £2,000 |
| Net contribution | £3,160 |
With a baseline of 115 a day the net contribution is £6,570; with 125 a day it is a loss of £250. Of the 1,200 extra sale orders, 970 came from returning customers and 230 from new ones, and returning customers account for most of the fortnight's dip. The shop records the sale as roughly break-even to modestly positive, mostly borrowed from its own regulars, and plans the next one as a test of a smaller discount with a holdout of email subscribers.
Common mistakes
- Judging the sale on its own days. The waiting before and the dip after belong to the result.
- Comparing with the week before only. One baseline, chosen after the fact, gives the answer people hoped for.
- Revenue instead of contribution. The discount comes entirely out of margin; revenue hides it.
- Ignoring the discount on orders that would have happened. It is a cost on every order the business would have had anyway.
- Counting every sale buyer as new. Customer status comes from the order system, and a sale's buyers may well be mostly existing customers.
- Assuming new customers are good customers. Follow their repeat purchases before crediting the sale with them.
- Leaving codes uncontrolled. Codes that spread beyond their channel turn targeted offers into general discounts.
- Reporting a single figure. Without a control group, the honest answer is a range.
The measurement planning hub lists the other guides that this analysis draws on.
