Case Study 2 — The Channel Trap: A Restaurant That Grew Delivery to 28% of Sales and Could Not Get Out
This is a labeled composite. The restaurant does not exist. Every operating pattern in it — the 2020 pivot, the ratcheting promotional spend, the ratings dependency, the throughput failure at peak, and the difficulty of unwinding a channel once guests have learned it — is drawn from well-documented industry experience in the years after the COVID-19 shutdowns. All figures are constructed for teaching and are internally consistent. No real business's financials are represented.
Background
A 96-seat modern American restaurant in a dense urban neighborhood of a large U.S. city. Opened 2016. Full bar, dinner six nights, weekend brunch, a check average of $52 at dinner, and roughly $2.3 million in annual sales by 2019. Two owner-partners: a chef and a general manager. No delivery, by choice — the chef's position, stated publicly more than once, was that the food did not travel.
Prime cost in 2019 ran 61%, which is a point over the full-service benchmark from Chapter 1 and entirely ordinary. Operating profit before debt service was about 7%. A good, unremarkable, well-liked independent.
What happened
Phase one: the emergency (2020)
The dining room closed by public order in March. Within eleven days the restaurant was on two marketplace platforms with a 34-item menu that was, essentially, the dinner menu with the prices copied across.
This was correct. It was correct even though almost every specific decision inside it was wrong. A restaurant with zero revenue and a payroll does not have the luxury of a tasting drive and a menu matrix; it has the luxury of being open. Chapter 28 §28.1 makes this point deliberately, and it applies here: the emergency decision was right, and the failure was never revisiting it.
By December 2020 off-premise was 71% of a much smaller business. The partners had learned to pack. The chef had stopped saying the food did not travel.
Phase two: the ratchet (2021–2022)
The dining room reopened. Off-premise did not go away — it settled at about 28% of a recovered $2.4 million in sales, roughly **$672,000 a year.**
Three things happened in this period, each individually reasonable.
The promotional spend ratcheted up. A competitor two blocks away started buying sponsored placement. The restaurant's order volume dipped. The general manager turned on sponsored listings and a standing "$6 off $35" promotion to recover it. Volume came back. Nobody turned either one off, because the moment it might have been safe to was never identifiable.
The menu never got edited. The 34 items published in April 2020 were still published in 2022, including a fried item and two composed salads that had never once arrived in acceptable condition. The restaurant's marketplace rating drifted down. Lower ratings meant worse placement, which meant lower volume, which produced more pressure to buy placement.
The channel colonized the peak. Orders arrived heaviest between 6:30 and 8:30 — the same window as the dining room. Nobody had ever considered closing the channel during service, because the channel had been born at a time when there was no service to protect.
Phase three: the arithmetic (2023)
The general manager finally built the analysis. Here is what a normal week looked like.
[Constructed teaching example.]
ONE WEEK OF MARKETPLACE VOLUME
Orders 312
Gross menu sales $12,920.00
Average order $41.41
Commission @ 25% ($3,230.00)
Promotion funding (restaurant share) ($646.00)
Sponsored placement ($388.00)
Error refunds charged back ($465.12)
─────────────────────────────────────────────────────────
NET REMITTANCE $8,190.88
Effective take rate 36.6%
Restaurant's own costs on 312 orders
Food cost @ 31% of menu sales $4,005.20
Packaging @ $2.05 x 312 $639.60
Packing labor: dedicated position,
38 hrs @ $21.00 all-in $798.00
─────────────────────────────────────────────────────────
CONTRIBUTION $2,748.08
per order $8.81
as % of menu sales 21.3%
Twenty-one percent. The same $12,920 sold in the dining room, at a blended COGS of 28% and after card fees and china, would have contributed roughly $8,700 — about 3.2 times as much.
Then she ran the throughput analysis, and it was worse.
The kitchen's constraint was a six-burner range and a single plancha, jointly capable of about 52 plates an hour sustained. Between 6:30 and 8:30 on Friday and Saturday, delivery was consuming an average of 9 plates an hour out of that 52 — roughly 17% of the peak-hour capacity of a restaurant that was quoting 50-minute waits on those nights and turning walk-ins away.
[Constructed teaching example.] At a $52 check and 28% blended COGS, a dine-in cover contributed $37.44. In the binding hour the kitchen produced about 52 plates to serve about 44 covers, so contribution per plate was roughly:
44 covers x $37.44 = $1,647.36 per hour
$1,647.36 / 52 plates = $31.68 per plate produced
A delivery order at $41.41 contributing $8.81 typically consumed two plates from that constraint — $4.41 of contribution per plate produced, against $31.68 the same equipment was earning for the dining room.
Every delivery plate produced during the peak cost the restaurant about $27.27 of contribution. At 9 plates an hour, two hours, two nights a week: roughly $982 a week, or $51,000 a year, destroyed inside a channel that reported $672,000 of sales.
The trap: why they could not simply stop
This is the part worth studying, and it is why this case is here rather than a straightforward "they did the math wrong" story. The general manager had the analysis. Unwinding took eighteen months, and here is what stood in the way.
Guests had learned the behavior. Roughly a third of the delivery volume came from households within six blocks who had, at some point in 2020, been dining-room regulars. Turning the channel off did not automatically convert them back to tables; it converted some of them to a competitor's delivery listing. Cannibalization, once it has happened, is not symmetrical — you can lose a dine-in guest to your own delivery channel faster than you can win them back by closing it.
The ratings were an asset with a maintenance cost. Three years of accumulated marketplace ratings and review volume were genuinely valuable, and pausing the channel risked placement that would be expensive to rebuild. This is the cost of the blackout experiment in §28.3, made concrete.
A person's job was attached to it. The dedicated packer position — $798 a week — existed because of the channel. Removing the channel meant removing the position or finding other work for a specific human being, and the partners were not willing to do the first casually.
The sales line had become part of how everyone described the business. A channel worth 28% of sales is not something you shrink quietly. Staff, vendors, and both partners had grown used to the top-line number, and the conversation about cutting it kept getting postponed because it always sounded like a conversation about shrinking. Revenue that everyone has grown used to acquires a constituency.
And the chef had been wrong once already, publicly, about whether the food traveled — which made "we're pulling out of delivery" a harder sentence to say than the arithmetic suggests. That is not a financial obstacle. It was, by the general manager's account, the real one.
What they eventually did
Not a shutdown. A re-specification — very close to what Chapter 28 recommends for Bellwether, arrived at eighteen months later and at considerable cost.
- Turned off the standing promotion and sponsored placement. Volume fell about 15%, and contribution rose immediately: 8.0 points of effective take rate recovered against a smaller base. This took ten minutes and should have happened in 2021.
- Cut the menu from 34 items to 14, dropping everything in the "Never" quadrant of Figure 28.5. The rating recovered over about four months and refunds fell by more than half.
- Closed the channel from 6:30 to 8:30 on Friday and Saturday. Delivery volume on those nights moved partly to the shoulders and partly away. Dine-in ticket times fell by an average of six minutes and the restaurant stopped quoting 50-minute waits it could not honor.
- Built first-party ordering and put a conversion card in every marketplace bag. Direct orders reached about 22% of off-premise volume in the first year.
- Kept the packer, reassigned to prep and expo support outside the packing window.
Two years on, off-premise was about 17% of sales at roughly double the contribution rate. The channel finally paid.
The lesson
A channel is easy to enter and hard to leave, and nobody prices the exit when they enter.
Everything in Chapter 28 that reads as fussy — the tasting drive, the nine-item list, the closed window on Friday and Saturday, the deliberate decision to defer a marketplace to a written year-two test — is an attempt to avoid this eighteen-month unwind. The cost of specifying the channel correctly on day one is one afternoon. The cost of specifying it wrong is a guest base that has learned a behavior, a rating you are afraid to risk, a position you created, and a revenue line with a constituency.
Three transferable rules:
- Re-examine every emergency decision on a schedule. A decision that was right under a state of emergency is not automatically right in year four, and nothing in your reporting will flag it. Diary it.
- Compute contribution per unit of the constraint, not per order. The channel looked like $672,000 of sales at 21% contribution — thin but positive. It was actually destroying $51,000 a year at the peak, and no report the restaurant received would ever have shown that.
- Turn off what you did not authorize. Promotions and sponsored placement were 8.0 points of take rate, enabled without a decision, removed in ten minutes. Chapter 34's authorization principle is not only about comps and voids.
Discussion questions
- The 2020 decision to launch delivery in eleven days with an uncosted 34-item menu was, this case argues, correct. Do you agree? What would have had to be true for it to be wrong?
- Compute the contribution per constrained plate for a channel in a restaurant you know or can model. What is the equivalent of the "plancha and six burners" in that kitchen, and who could tell you its sustainable rate?
- Rank the five obstacles to unwinding by how hard each would be to overcome. Which one is a genuine business constraint and which are psychological? Does the distinction change what you would do?
- The restaurant recovered 8.0 points of take rate in ten minutes by turning off promotions and placement. Why did it take two years? Design the control that would have surfaced it in month one — who reviews what, how often, and on what report.
- Cannibalization is described here as asymmetric: a dine-in guest lost to your own delivery channel is harder to win back than to lose. Is that claim plausible? What evidence would you look for, and what would you do differently if you believed it?
- The contested decision: the partners kept the packer's job and reassigned the position rather than eliminating it, at a real cost to the channel's economics. Was that the right call? Argue it from the P&L, then argue it from Chapter 21's retention arithmetic, and say which argument you find more persuasive and why.