Repeat-purchase programs: the mistakes that bleed margin and the method that holds in 2026

Chapter 1 — The incentive comes out of contribution margin, not menu price

Giving away a 12 USD dish at 30% food cost costs you 3.60 USD in product and sacrifices 8.40 USD of contribution, and neither figure is the one most operators book. Whoever charges 12 USD to the marketing budget is reporting a cost that never happened, because that money was never in the till; whoever charges 3.60 USD ignores the margin forfeited at a table that was in fact occupied. The number you decide with is the sacrificed margin, and it is the one that almost never reaches the P&L. With customer acquisition cost up 222% over the eight years through 2025 (Marqii 2025), getting the costing base wrong by a factor of three or four stops being an accounting detail and becomes the difference between a program that funds growth and a spending line nobody can defend. Without a deliberately withheld control group, you cannot tell whether you bought frequency or subsidized guests who were returning anyway.

Chapter 2 — How do you know the program produced the visit rather than paying for one already coming?

Set aside 5% to 10% of the base, leave it without access to the benefit for two full quarters, then compare frequency and check against the exposed group:

that difference is the only thing you may call incremental. Paytronix (2024) reports customer value rises 23% with rewards programs, but that 23% describes exposed populations, not a net lift in your operation. Diego F. Parra presses this point because the mistake recurs in every Masterestaurant review: an 18% jump in member visits gets celebrated when it lands squarely on high season, and the program takes credit that belongs to the calendar. The control group costs a few points of redemption and saves the entire budget. In an operation below 500 thousand USD in annual sales, any platform with a fixed monthly fee eats the margin before producing a single extra visit.

Chapter 3 — Under 500 thousand USD a year: run it by hand before you buy an app

At 480 thousand USD in revenue, a reasonable prime cost and operating profit between 6% and 10%, you are looking at 28,800 to 48,000 USD of yearly EBITDA: a 350 USD monthly license is 4,200 USD, somewhere between 9% and 15% of all profit, committed before anything gets measured. The recommendation for this band does not change and it is no consolation prize: your own email and messaging, a list segmented by frequency, and an incentive charged against the dish with the best margin. Omnisend (2024) measured a 25.1% average email open rate, and Stripo (2025) reports 26% higher opens with personalized messages; across a base of 1,500 guests that means hundreds of monthly touches at near-zero marginal cost. Platforms begin to justify themselves above 1 million USD in annual sales, not before, and the arithmetic is uncomfortably plain. Between 500 thousand and 1 million, that same 4,200 USD license represents 5% to 8% of EBITDA and demands incremental frequency that few bases under 3,000 guests produce.

Chapter 4 — From 500 thousand to 1 million, and 1 to 5 million: where the platform starts to pay

Past the million mark, and especially in the 1 to 5 million band where the three-to-ten-location group lives, the fixed cost dilutes below 1.5% of sales and consolidated cross-site data starts earning its keep. A four-location group at 3.2 million in sales moving average frequency from 2.1 to 2.4 annual visits across 12,000 active guests generates 3,600 extra visits; at a 22 USD check and 62% contribution margin, that is 49,100 USD against 50,400 USD of license plus incentives. Breaking even is not winning. Once the operation clears 5 million USD and, all the more, 10, the incremental visit you bought carries labor and capacity costs that smaller bands could safely ignore. The National Restaurant Association (2024) documented a 35% rise in labor costs since 2019, and 7shifts (2024) measured base pay at 14.20 USD per hour after a 4% increase: pulling 8,000 extra visits into a location already running at 85% occupancy during peak does not fill empty tables, it forces another shift open.

Chapter 5 — Above 5 and 10 million: the cost of serving the extra visit stops being zero

There the correct calculation subtracts three things from incremental margin — incentive, platform and extra service hours — and the result frequently turns negative precisely in the daypart where the program reports its best performance. The fix is not switching the program off, it is dating the incentive toward Tuesday and Wednesday, where idle capacity makes the extra visit cost barely more than its ingredients. A celebrity restaurant or a large-format themed venue above 5 million USD does not buy repeat purchase: it buys demand distribution. Its problem is not that guests fail to return — with a waiting list, repeat business happens on its own — but that acquisition cost is already paid by the name and by outside creators, with bookings up 30% the week after a creator posts (Marketing LTB 2025). Giving product away there destroys value perception and competes against its own waiting list. What does work at that tier is a non-discountable benefit: early access to menu changes, the chef's table, a seating slot ahead of general opening.

Chapter 6 — The celebrity-chef and large-format themed restaurant play with different numbers

It costs valley-hour capacity rather than margin, and it holds price. One additional review star moves revenue between 5% and 9% (Harvard Business School, Michael Luca), and that channel returns more than any coupon. A repeat-purchase program is profitable when incremental contribution margin exceeds the sum of three costs, and that comparison fits on one line: margin per additional visit, minus the incentive sacrificed, minus the prorated platform, minus the labor cost of serving. Toast (2026) reports that 33% of industry professionals name attracting and retaining guests as their top challenge, ahead of anything else, and that consensus is exactly what makes the spending dangerous: nobody argues with a budget everyone considers a priority. At Masterestaurant the test we require before renewing a license is two quarters with an active control group and costing based on contribution. If the incremental delta fails to reach 1.8 times the program's total cost, it does not get renewed and the money goes back into menu and service.

Chapter 7 — The profitability threshold that belongs on the committee's first slide

Run that ratio on last closed quarter's numbers today, before you sign the next annual contract. The first is pure arithmetic and explains most of the rest: costing at menu price versus costing at contribution margin shifts reported spend by a factor of three or four. At 30% food cost, a free 12 USD dish consumes 3.60 USD of inputs and sacrifices 8.40 USD of contribution; the operator logging 12 USD is reporting a cost that never existed, and the one logging 3.60 USD is ignoring margin never collected. The decision-grade number is margin sacrificed, and almost nobody tracks it. The second is methodological: without a control group you do not have a program, you have decoration. Unless you deliberately withhold 5% to 10% of your base from the benefit for at least two quarters, you cannot separate the visit your incentive caused from the one already on the guest's calendar.

Chapter 8 — Five differences that decide whether the program funds EBITDA or eats it

It is uncomfortable, and someone on the committee always wants it gone by month two; it is also the only line that turns a marketing report into a financial document a board can defend. The third lives in the menu and few people see it: which dish you give away matters more than how much you give away. Menu engineering sorts dishes by popularity and contribution margin, and the incentive belongs in the high-margin, low-popularity quadrant — the puzzle — never on the star. Discounting the star cannibalizes sales you already had and sacrifices the highest margin in the house; discounting the puzzle buys rotation for a dish that was occupying menu space without paying rent. The fourth is frequency segmentation, where commercial instinct usually gets it backwards. The high-frequency guest who already comes three times a month gets no discount: they get access, recognition and priority, which cost close to nothing in inputs and buy real loyalty.

Chapter 9 — Five differences that decide whether the program funds EBITDA or eats it — in practice

Monetary incentive, which does burn margin, is reserved for the dormant segment with no visit in 90 days, where sacrificed margin actually purchases a trip that was not going to happen. A flat program subsidizes precisely the guest who needs no subsidy. The fifth is horizon, and it settles most arguments: the verdict on a repeat-purchase program arrives in month six, not month one. Month one the whole base redeems accumulated balances and spend spikes while incremental revenue lags; by month three the noise drops; by month six the frequency curve by cohort reads cleanly against control. Paytronix (2024) puts value per customer 23% higher where a well-built rewards program runs, and that 23% is a cumulative effect no monthly close can show.