Acquisition cost down 5.8 points: how we fixed a customer retention program that lived in the brochure and not in the till, using Radar Gastronomico and meseros.ai

The number nobody had calculated in seven years of operation

Seven years of trading and nobody had run the one number that governed the business: 71% of the people who ate there in the baseline month were walking in for the first time, and only 14% came back inside ninety days. Put those two figures side by side and the rest explains itself, including why digital ad spend climbed 31% year over year while sales moved 9%. They came out of a manual cross-check, tickets against booking phone numbers, three afternoons of work no installed tool did on its own. Everything else looked orderly: monthly accounts current, declared food cost near 30%, a 104-seat room with its own kitchen and 31 people on payroll, average checks of 11.40 USD in the dining room against 14.60 USD on digital orders. On paper, no fire. At the exit door, plenty.

Why can a profitable restaurant be going broke through its funnel?

Because the funnel eats the cash: whoever repurchases an entire clientele every quarter is funding permanent commercial spending out of working capital, and accounting profit never shows it.

Between 1,900 and 2,100 distinct guests crossed that door each month, and at 14% ninety-day recurrence nearly the whole base had to be paid for again. The arithmetic of the wheel is harsh. Winning a stranger runs considerably dearer than bringing a guest back, especially with lower prices already acting as a visit driver for half of diners, according to Circana (2025). Multiply 2,000 guests by 30 dollars, set the product beside the EBITDA of a 104-seat location, and you can see it never closed.

Third-party delivery was not a channel, it was a hemorrhage with a logo

With 38% of orders arriving through outside platforms, margin evaporated twice on the same trip: first at the commission, then at the uncosted promotion. Service fees and mandatory mechanics together left 30% to 40% of order value in the platform's hands (Restaurant Business, 2024), and the algorithm rewarded precisely what a consolidated P&L was unable to separate. Costing dish by dish exposed the gap, 5.9 points between the 30% theoretical cost and what actually crossed the till. The same plate returned 58% contribution margin served at a table and 41% when it went out promoted. Seventeen points handed away for a diner whose name, phone and frequency never reached the operation. The real price was never the commission. It was the BLINDNESS.

How recurrence was instrumented with Radar Gastronómico?

No points program at the start: measurement first.

Radar Gastronómico from Masterestaurant instrumented identity capture at the three contact points, booking, table and digital order, and the meseros.ai dashboard started returning weekly cohorts for first visit, second visit and ninety-day frequency, the metric that rules every other one. The Restaurant Model Canvas redrew the value proposition channel by channel; the cash-flow module moved commercial spend to where it belongs, an investment with returns read by cohort rather than a fixed expense line. What landed first was not a campaign but a single-screen board carrying three numbers: unique diners for the month, second-visit rate, cost of demand replacement. As Diego F. Parra, restaurant consultant and founder of Masterestaurant, puts it in every diagnosis: whatever has no cohort has no owner.

What was done with the data once it existed?

The cheapest lever, once identity existed in the base, turned out to be direct messaging, and the channel evidence leaves little room to argue:

SMS marketing converts between 21% and 30% on average (Constant Contact, 2024), a range no paid media comes near. We built a two-touch sequence at seven and twenty-one days after a first visit, anchored to the highest-margin dining-room dish instead of the one the platform pushes. Social content shifted toward footage the guests shot themselves, which converts 4 times better than brand photography (Loop.fans, 2025) and whose posts perform more than 10 times above those without it (Emplifi, Q3 2025). No photographer, no agency: a happy customer's camera and a release signed at the table.

The measurable result and the tension nobody wants to name

Ninety-day recurrence climbed from 14% to 27% across two quarters while ad spend fell 22% in absolute terms and sales kept rising, according to the case's cohort tracking. Then came the paradox that frightens any owner: cutting the ads made the monthly count of unique diners FALL, from around 2,000 to roughly 1,750, and that count was exactly what his dashboard had celebrated for seven years. Sales rose anyway. A guest who returns three times is worth more than three guests who come once and drag CAC behind them; the bridge between both readings is margin per cohort, not headcount. I got this wrong for years by recommending the points program first: with no second-visit measurement, a program is a discount with a pretty card.

Transferable lessons

What you do this week depends on your revenue band, and the order is not negotiable. Under 500 thousand dollars a year, ask for a phone number at every table for fourteen days and count the repeats by hand, on paper, with no software involved. In the 500 thousand to 1 million band, the one described here, instrument capture at the three contact points and publish the ninety-day second-visit rate on the weekly board. Past a million, cross cohorts against contribution margin by channel and stop paying for platform promotions that land under 50% margin. Above 5 million there is room for someone accountable for recurrence, with a budget and a bonus tied to second visits. And for multi-site groups past 10 million, consolidate one guest identifier across locations before buying any CRM.

Limits of this case

Three contexts where I would not expect these numbers, and saying so beats watching somebody copy the recipe. Pure-transit operations, airports, stations, tourist zones with total turnover, have the ninety-day second visit structurally closed, and chasing it burns cash: the lever there is average check and table velocity. Bureau of Labor Statistics), since bringing a guest back to a dish that loses money only speeds up the collapse. Fix the costing first. The third case is inconsistent product: recurrence amplifies whatever exists already, so if the experience fails one visit in three, measuring it documents the leak with better precision and nothing more. Real patterns were merged under anonymity here, with no result promised.

Root-cause diagnosis: every symptom and the number that exposed it

SYMPTOM: ad spending grew 31% year over year while sales grew 9%. ROOT CAUSE: the operation bought traffic without capturing identity, so the base had to be replenished whole every quarter. WHAT EXPOSED IT: 71% first visits in the baseline month, a figure no internal tool produced, pulled from matching tickets against booking phone numbers. SYMPTOM: declared food cost read 30% and the till read something else. ROOT CAUSE: uncosted platform promotions, the buy-one-get-one kind an algorithm rewards and a consolidated P&L cannot separate, opened 5.9 points between theoretical and actual cost. THE PROOF: 41% contribution margin on a promoted dish against 58% on the same dish served at a table. SYMPTOM: guests called themselves satisfied and still did not return. ROOT CAUSE: satisfaction and recurrence are different variables and only the first was measured; at 94% floor turnover, the month's fourth visitor was served by someone three weeks into the job.

Root-cause diagnosis: every symptom and the number that exposed it — in practice

THE NUMBER: 4.6 out of 5 in public reviews sitting next to 14% ninety-day recurrence. SYMPTOM: the treasury tightened in the third week of every month, without fail. ROOT CAUSE: a deferred P&L hid that commercial outlays left upfront while the sales meant to cover them arrived twenty-eight days later through platform settlements. WHERE IT SHOWED: 19 days of working capital locked in delivery receivables, against 3 days on the owned channel. SYMPTOM: the owner spoke of his clientele by first name and the system knew none of them. ROOT CAUSE: guest knowledge lived in two memories, his own and a host's with six years in the house, and memory neither scales nor survives a resignation; that is a liability, not an asset. EVIDENCE IN THE TILL: 340 usable exportable records for an operation serving more than 24,000 guests a year.