Delivery algorithm optimization: the method that burns margin vs the one that wins the ranking

What actually moves the ranking: the discount or the kitchen clock?

The kitchen clock moves the ranking; the discount only moves that week's volume. An average customer scrolls 2.4 screens before choosing, and across that sweep the aggregator sorts by estimated reliability rather than generosity:

acceptance time under 60 seconds, real prep time matching the promised one, cancellations below 2%. On the other side sits the 20% promo, which lifts orders during the nine days of the campaign and drops them back on the tenth, because it bought borrowed demand from people who already knew you. Persistence is where the gap shows: the operational signal builds a history, the promotional one expires with the budget. With more than 40% of adults ordering delivery three to five times a month (UpMenu, Food Delivery Statistics 2024), the pool of new customers is huge and you are paying for it with price instead of earning it with minutes. The clock wins. A USD 14 ticket at 28% commission leaves USD 10.08 before ingredients; apply a 25% promo and the aggregator charges its commission on the final USD 10.50, so USD 7.56 reaches your till.

Discount arithmetic versus minute arithmetic

That USD 2.52 drop per order, across 400 monthly orders, is USD 1,008 spent to win customers who mostly belonged to you already. Now look at the other column: cutting acceptance time from 180 to 45 seconds costs zero margin, it just requires a tablet with sound and someone accountable for touching it. And the effect does not wear off. Diego F. Parra keeps insisting at Masterestaurant that owners price the implicit acquisition cost of every discount point before signing it, because the promotions panel never displays that subtraction. Once margin per order stops covering the packaging, you no longer have a marketing problem. Declaring 20 minutes of prep and delivering in 34 is the most expensive lie available inside the app. The courier arrives at minute 20, waits 14, and that wait is logged against your restaurant rather than the rider; food goes out lukewarm, the rating slides from 4.7 to 4.4, and the system starts showing you lower exactly during peak traffic.

Promised time versus executed time: the metric that punishes quietly

The uncomfortable alternative, declaring an honest 30 minutes, shrinks short-term volume because you look slower than the place next door, yet it stabilizes your history within a few weeks. My position here is firm: give me a slow, accurate restaurant over a fast, dishonest one, since the first accumulates reliability and the second accumulates debt. What would happen if you held 34 honest minutes for a full quarter? The algorithm would learn your truth and stop ambushing you every Friday night. Flagging items out of stock at ten at night destroys more position than running a short menu. Every block counts as a broken promise, and a 60-dish catalog with 11 switched off on a Friday projects an 82% availability rate that the system reads as risk. Facing it is the curated route: 22 dishes that always ship, with your own photography and descriptions of 15 to 25 words, holding 99% availability and steady prep because the kitchen repeats movements instead of improvising them.

Sold-out catalog versus curated catalog: two ways to lose visibility

The operating math shifts too, given that fewer references mean less waste and fewer dead purchases sitting in the walk-in. I got this wrong for years while recommending menu breadth as a ticket lever; delivery data says the opposite, and inside a dark kitchen catalog depth is a liability. The short catalog wins. A dark kitchen running two brands from one site was spending USD 1,450 monthly on funded promotions and billing USD 21,300 across 1,520 orders. We killed every campaign for six weeks and attacked three numbers: acceptance from 210 to 38 seconds with a dedicated tablet, declared prep from 18 to 27 minutes, which is what the kitchen truly executed, and cancellations from 4.1% to 0.9% after trimming the catalog from 54 to 24 references. Volume fell to 1,190 orders in week two, and that is the frightening part, the one that makes most owners abandon the experiment.

The dark kitchen that switched its promotions off

By week six orders reached 1,605, the rating climbed from 4.3 to 4.8, and monthly contribution improved by USD 2,900, because those 1,605 orders came in at full price. The ranking did not reward the discount: it rewarded consistency. Forecasting demand by time slot is worth more than any campaign when your real problem is the 8 p.m. bottleneck. Scheduling tools built on artificial intelligence report labor cost reductions of 8% to 12% with forecast accuracy above 90% (TimeForge 2025), and in delivery that forecast converts straight into minutes: two extra hands between 7:30 and 9:30 p.m. hold your declared prep time through the peak, which is precisely when the algorithm is watching. Comparing that against a promotion is brutal, because the promo floods the same slot with orders your kitchen cannot execute and multiplies the damage. Meanwhile, each staff departure costs up to 150% of the salary in replacement (StaffedUp 2025), so sizing the shift correctly protects both the operation and payroll.

What scheduling AI does for your prep time?

Schedule first, promote afterward, never the other way around. Promotion works in two concrete scenarios and nowhere else:

the launch of a virtual brand with no history, where no operational data exists for the system to read, and recovery after a specific rating drop already fixed in the kitchen. Both cases mean windows of 10 to 14 days with a defined spending cap, not a permanent 20% discount that customers absorb as your list price. The tension is genuine, because without initial visibility there are no orders and without orders there is no history to optimize; it resolves by flipping the sequence, since you tune the times with whatever volume you have and buy the push afterward. A restaurant promoting on top of 34 real minutes and 4% cancellations is paying to show its worst face to more people. That is the expensive mistake.

What to choose according to your operating profile?

If you run fewer than five sites or a dark kitchen, sort your times out and forget the promotions panel for eight weeks:

acceptance under 60 seconds, declared prep equal to executed prep even when raising it from 18 to 27 minutes stings, cancellations below 2%, catalog availability above 97%. That package costs no margin and produces accumulated position. If you operate a chain with more than fifteen points and already hold those four numbers, then tactical promotion does pay, because you are buying reach on top of an operation that absorbs the hit. And if your restaurant sits in Colombia, where menu prices rose 9.8% since February 2025 to sustain 98,000 jobs (ACODRES 2025), giving away 25 points of price is a decision you justify with numbers, not with fear. Measure your acceptance time tonight. The discount buys DEMAND; the operation buys POSITION. Two different markets, and that is why owners get confused: they watch orders climb with the promo and conclude the algorithm rewarded them, when what happened was that their own customers bought cheaper for nine days.

Where the comparison breaks?

Aggregator commission is charged on the final discounted price, so a 25% promo does not cost you 25 points:

it costs 25 points of price plus the commission you still pay on the rest, and on a USD 14 ticket at 28% commission that leaves a contribution that sometimes fails to cover packaging. Operational signals compound and promotional ones do not. Every on-time order feeds a history the system uses to estimate your reliability; every campaign, by contrast, is judged inside its own window and forgotten when it closes. There is a paradox worth resolving before going further: the engine needs your data to show you, and you need to be shown to generate data. The way out is not discounting the whole catalog, it is discounting ONE dish until it accumulates enough sales history and ratings, then switching off. You are buying information, not volume.

Where the comparison breaks — in practice?

Catalog weight surprises almost everyone: an item without a photo converts far below one with professional photography, and in a ghost kitchen running a virtual brand, where the customer has no storefront to look at, the photo IS the restaurant.

Running a dark kitchen from scratch multiplies the advantage of the operational method, because you carry no inherited dining-room menu and can design those 20 items around cook times, packaging and travel from day one.