Delivery app algorithm: how a dark kitchen cut cancellations and brought delivery EBITDA back to positive in 90 days with the Standard Recipe Generator

What does the delivery app algorithm really measure?

The delivery app algorithm measures whether the order will arrive well, and advertising only buys a shop window that operations then have to sustain.

Platforms rank the list with fulfillment signals: how many orders you accept, how many you cancel on your own account, how your rating has trended over recent weeks, how far real time drifts from promised time and how many listing visits end in a purchase. That shop window weighs more where a single app concentrates the market, because according to Sensor Tower (2024) iFood held 89 % of monthly active delivery users in Brazil, so dropping in its listing means vanishing from almost all of the channel's demand. At Masterestaurant we explain it with a kitchen image: the app is an expediter who never tastes the food, only checks whether it leaves on time, complete and without returns, and punishes whoever fails it two nights in a row.

Starting point: a ghost kitchen losing rank without knowing why

The kitchen in this case lost visibility because of production failures and not for lack of ad spend, a diagnosis that changed the whole work plan. It is a ghost kitchen running three brands, sushi, bowls and burgers, which in this illustrative composite case sold almost everything through apps and watched its listings slide on weekends. The partners had already priced paid campaigns, convinced the algorithm punishes whoever does not invest. The platform dashboard told another story, because the drop matched the nights with the most rejections and the week the sushi brand began cancelling over salmon stockouts. Buying ads at that moment would have meant sending more customers to a kitchen that could no longer deliver for them, and every failed order would have sunk the listing a little further. First you fix the kitchen, then you buy visibility: that ORDER is the thesis of the case.

Why does a tablet with no owner sink the acceptance rate?

An order tablet with no one responsible per shift sinks acceptance precisely in the hours that matter most, because nobody looks at it when the kitchen is slammed.

In this kitchen the device lived on the pass and was handled, in theory, by the same person plating, so rejections and expired orders piled up on the two busiest nights. To the algorithm, every order that expires unanswered signals that the location is unavailable, and the listing drops for every customer in the area, not just the one who was waiting. The fix was organizational, not technological: a named tablet owner per shift, a written handover for the rush and one simple rule, pause a brand rather than let its orders expire. Pausing hurts that hour's sales. Rejecting hurts the whole week's ranking, and that bill arrives later and without warning.

Salmon as the cause: cancellations the customer had already paid for

The cancellations in this case had an ingredient's name, salmon, and they stopped once inventory started talking to the app. The stockout was never flagged on the platform, so the customer ordered sushi, the kitchen discovered the shortage while assembling the order and cancelled after the charge had gone through, the worst possible combination for rating and ranking. That is where the Masterestaurant method's recipe costing tool came in: with each roll's costed recipe we calculated how much salmon the forecast sales per time slot consumed, set a reorder point and defined who marks the item as sold out in the app before service opens. Skipping that is expensive. In Mexico, the president of CANIRAC estimated that delivery apps charge between 15 % and 35 % of sales (El Universal San Luis, 2026), and with that commission load every cancelled order is committed margin walking out the door and reputation lost for nothing.

Promised time versus real time on two-brand orders

The prep time configured in the app was the ideal-recipe time, and the algorithm charges for every minute of gap between what was promised and what happened. The deviation grew on combined orders, because bowls and burgers came from different stations and nobody coordinated closing the bag, so the courier waited, the food arrived lukewarm and the review said so. Diego F. Parra argues for an order that seems counterintuitive: raise the promised time to what the kitchen truly delivers at peak and only then trim it with method. A longer promise loses the occasional click; a broken promise loses rating, and rating moves the list for weeks. The operational answer was an assembly lead for two-brand orders, with a ticket board showing which station finishes last. What would have happened if they had cut the time to compete? More late orders, worse reviews, lower rank, and a cycle that feeds itself.

Result: the listing climbed back before any ad spend

Ranking recovered once the kitchen was in order and before spending a single dollar on ads, the result that most annoys anyone selling ad space. In this illustrative case, by the third month acceptance came close to every order received, cancellations caused by the kitchen became rare exceptions and the delivery channel stopped dragging down the income statement. Only then did the team work on listing conversion, with real photos of each bowl, dish names customers actually search for and a two-brand combo priced so it does not cannibalize the others. Price matters because customers compare the total they see on screen, and in the Intouch Insight (2025) report third-party app fees in the US fell by an average of 1.10 USD between 2024 and 2025. With advertising on top of an operation that delivers, every purchased visit finally has a reasonable chance of ending in a well-delivered order.

Transferable lessons by annual revenue band

The lesson travels in the same order at any size: first the operation the algorithm measures, then the visibility you buy. Under half a million dollars a year, this week's first step is to name a tablet owner per shift and write down the handover. Between half a million and one million, this case's band, the move is to cost the ingredient that causes the most cancellations and give it a reorder point. Above one million, each brand needs its promised time measured at peak, not the recipe time. Past five million, as with a large-format themed concept or a celebrity chef's brand, the step is to audit which app dominates each city, because in Mexico DiDi Food held 38 % of monthly active delivery users (Sensor Tower, 2024). And a group above ten million should unify the cancellation dashboard across all its locations before sitting down to negotiate commissions.

Limits of this case

This result would not repeat the same way in every market, and that is worth saying before copying the method. Where a single app gathers nearly all demand, as in Argentina, where PedidosYa held 61 % of monthly active delivery users according to the same Sensor Tower tracking from 2024, the kitchen depends on one algorithm and any rule change weighs more than operational gains. Nor would I expect the same where the platform map is shifting: Uber Eats went from 33 % in 2020 to 7 % of Latin American users in 2024 after leaving Brazil, and a well-ranked listing can end up stuck in an app that is losing customers. A restaurant with a dining room, for which delivery is a side channel, hits a different limit, the kitchen shared between floor and bag, where the tablet competes with the tables. There the first job is to separate stations, not to polish the listing.

Root-cause diagnosis, transferable lessons and the limits of the case

Behind every symptom sat a cause that had nothing to do with marketing. Low acceptance came from a tablet with no owner at the pass, and the giveaway was that rejections clustered on the two busiest nights, exactly when the pass cook was plating. Cancellations had an ingredient's name, salmon, whose outages were never flagged in the app, and late deliveries grew on orders mixing two brands from different stations with nobody closing the bag. I got this wrong for years, recommending listing work and ads first because they are visible, when the delivery app algorithm is really an accountant of kept promises that won't believe a pretty photo if the kitchen fails. Had these owners bought ads in week 1, more orders would have landed on the same ownerless tablet, rejections and cancellations would have risen, organic rank would have dropped further, and they would have paid twice, in ads and in commission, for customers who never came back.

Root-cause diagnosis, transferable lessons and the limits of the case — in practice

The channel's tension is real: the app that charges the commission is the one that brings the customers. The way out is to treat it as paid acquisition with a ceiling. Market concentration makes that urgent; one platform held 33 % of Latin American monthly users in 2020 and fell to 7 % in 2024 after leaving Brazil (Sensor Tower, 2024), so rank won inside one app is rent, while standard recipes, stopwatch times and a customer base of your own work anywhere. Lessons scale by size. Under half a million dollars a year, the first step is naming a tablet owner per shift this week. Between half a million and one million, like this case, it is locking portions and real times for the brand that cancels most.

Root-cause diagnosis, transferable lessons and the limits of the case — key points

Above one million, it is a single dashboard of the five signals per brand and location. For groups above five or ten million, including the celebrity-chef venue and the large themed experience restaurant, image royalties and set upkeep belong in break-even, never on the plate, and the first step is a short travel-proof delivery menu; off-premises traffic in US full-service reached 30 % in 2024, up from 19 % in 2019, per the National Restaurant Association. Limits, to avoid survivorship bias: I would not expect the same result where one app holds nearly everything, as in Brazil, where the leader had 89 % of monthly users in 2024 according to Sensor Tower, because there negotiation and visibility outweigh operations. Nor in a kitchen located too far from its demand, nor with a dish that simply doesn't survive the ride.