Delivery app data and reviews: the playbook that separates before from after

Step 1: pull the last 40 orders and forget the lifetime average

Download your last 30 days of order data and isolate the most recent 40 rated deliveries, because that short window is what moves your ranking inside the app while the cumulative average exists mainly to comfort the owner. A restaurant sitting on 1,200 reviews and a 4.6 lifetime score can be averaging 3.9 across its last 40 deliveries without the headline number budging a single decimal for six or seven weeks, and by the time the visible rating slides to 4.4 you have already lost two tiers of exposure. What this step delivers is a spreadsheet with four columns per order: exact time, dish, rating, and the customer's actual words. Nothing else. Verify it by counting rows: if you have fewer than 40 rated orders, widen the range to 60 days before drawing any conclusion, since below that floor noise beats signal every time. Sort those 40 rows by clock time and average each sixty-minute block, which is where the real pattern surfaces.

Step 2: cross rating against dispatch time to find the hour that is bleeding you

At the Medellín fast-food operation billing 41 million pesos a month through Rappi, the slide from 4.6 to 4.1 was not evenly spread: between 19:40 and 21:10 the average sat at 3.4 while the rest of the day held at 4.7, so the problem was never the kitchen but the kitchen AT PEAK. Twelve million pesos gone in two months over ninety badly handled minutes. Your deliverable here is a single bar per time slot, drawn by hand if that is faster. You will know it worked when you can point a finger at the guilty window and state how many orders it represents against the total, which typically lands somewhere near 30%. Any one-star review naming a dish, an hour, or a temperature is worth more than twenty silent five-star ratings, because it hands you free the audit a consultant would bill for.

Step 3: read the bad reviews hunting for the data point, not the insult

When a customer writes that the burger arrived cold at 21:40, that person is describing with surgical precision an assembly-time or packaging failure on the hot line during peak. The mistake I see over and over is answering that comment defensively and filing the person away as difficult. Tag every negative comment into one of four buckets: temperature, missing items, delay, food quality. Then tabulate. The deliverable is a plain count, and across most operations I review two buckets hold more than 70% of the complaints, which shrinks the whole problem down to two attackable causes instead of a vague feeling that service has gone soft. When temperature dominates, the fix almost never lives in the food and almost always lives in assembly order and packaging. Move the fries into a vented container, hold the bag build until the courier is under three minutes away, and raise your auto-accept threshold so the line never carries six simultaneous tickets it cannot sustain.

Step 4: attack the dominant cause with a measurable operational change, never a discount

Diego F. Parra keeps hammering at Masterestaurant that the prep time you declare in the app is a contract with the algorithm: declare 18 minutes and dispatch at 27, and the platform punishes your on-time rate long before any customer punishes your stars. Raise the declared time to whatever you actually hit. Losing three positions for promising speed costs more than showing up as an honest 25-minute kitchen, and 76% of US operators already treat technology as a competitive edge (National Restaurant Association, 2024) precisely because these adjustments can be measured. Ghost kitchens play by harsher rules, since they have no storefront, no server rescuing a mediocre meal with a smile, and no neighborhood passing the name around. Their only trust asset is the app listing: stars, photos, comments. A ghost kitchen at 4.3 is effectively closed half the day even with the lights on and the full crew clocked in.

Step 5: why a dark kitchen cannot afford a single tenth of a star

Flip it around for a second: if Rappi hid ratings tomorrow, the brick-and-mortar restaurant would keep selling through its window and word of mouth, while the dark kitchen would be reduced to fighting on price against something like 74,000 listings, roughly the universe DiDi Food runs in Mexico, where 70% are local small businesses (DiDi Food, 2024). That asymmetry explains why inside a ghost kitchen every tenth of a star gets defended like margin. The first and costliest is buying reviews or begging relatives for them: fraud systems at Uber Eats, which pushed close to 74.6 billion dollars in gross bookings during 2024 (Statista, 2024), spot clusters of fresh accounts with no order history and void the entire batch, leaving you worse off than when you began. Second, staring only at the aggregate score. Third, answering every comment with the same copied template, which any customer smells in two seconds.

Five mistakes that wreck this work before it starts

Fourth, switching the app off during a crisis believing that protects the score, when inactivity hurts placement as much as a bad star or more. Fifth, changing five things at once: move packaging, declared time, menu, and price in the same week and you will never know which one worked. One change, two weeks of measurement, then decide on data. Mark your start date and review the board every fourteen days with four numbers on the table. Recent-window rating should climb at least two tenths per fortnight until it settles near 4.7. Complaints in your dominant bucket should fall below 20% of all negative comments. Actual dispatch time in the critical slot should converge on the declared time with under four minutes of drift. And your own cancellation rate should sit below 2%. If three of those four move the right way by week four, you are on track and the job becomes maintenance.

How to know it all landed: the eight-week close?

If none of them moves, your time-slot diagnosis was wrong and you go back to step two with 60 orders instead of 40.

Global online food delivery is worth 1.51 trillion dollars in 2026 and grows at 6.24% a year (Statista, 2026); inside that tide, your spot on the first screen is the only thing actually up for negotiation. Delivery app data and reviews are worked on the LAST 30 to 100 orders, never on the full history. Rappi and Uber Eats weight the recent window because they want to know how you cook today, not how you cooked in March. A venue with 1,200 lifetime reviews and a 4.6 average can be falling off a cliff while its last 40 orders average 3.9, and the big number hides it for weeks. A negative review carrying a concrete detail beats ten loose stars, because it tells you exactly what to fix.

Four differences that decide the outcome

When a customer writes that the burger arrived cold at 9:40 p.m., they are handing you a free audit of your hot line at peak. The mistake I keep running into is treating that comment as a personal attack instead of the cheapest operational input available. Replying does not raise the rating; fixing does. The public reply works on the future reader deciding between your listing and the competitor next to it, and there it does move conversion: a specific, non-defensive answer turns a two-star review into a signal of a serious operator. Leave the cause sitting in the kitchen, though, and the next batch of reviews repeats the same text. In a dark kitchen the effect hits harder than in a brick and mortar restaurant. A hidden kitchen lives off in-app ranking with no street traffic to cushion it, so half a rating point translates straight into lost orders; a venue with a dining room absorbs that hit through reservations and walk-ins while it repairs the digital channel.