EBITDA from -1.9% to 8.6%: fixing own channel vs delivery apps in a 3-brand dark kitchen with the Restaurant Model Canvas and the Demand Radar
Own channel vs delivery apps is not an either-or decision, and whoever frames it that way has already lost: apps buy DISCOVERY at an intermediation cost of 22% to 30% of the ticket, while the own channel buys REPEAT PURCHASE at an acquisition cost that falls with every returning order. This operation billed 780,000 USD a year with 94% of volume inside the platforms and an EBITDA of -1.9%; after seven months of rebalancing the mix settled at 61% apps and 39% own channel, Prime Cost dropped from 71.4% to 63.8% and EBITDA closed at 8.6%. The rule I leave written down, because everything else rests on it: treat the apps as your FIRST-PURCHASE channel and measure their profitability against the cost of equivalent geotargeted advertising, never against zero.
The case file first, interpretation later. A dark kitchen running THREE virtual brands out of a single 96-square-meter kitchen in a mid-sized Colombian city, 11 employees across production and dispatch, a revenue band of 500,000 to 1 million USD a year (780,000 closed the year before the intervention), a 9.40 USD average ticket, four years in business and one brutally dominant channel: 94% of orders arrived through Rappi, Uber Eats and DiDi Food, with the remaining 6% split between WhatsApp and an ordering page the owner had built two years earlier and never promoted.
The owner arrived with a sentence I hear almost verbatim several times a year: revenue looked fine, but the money evaporated inside the platforms. He was half right, because the commissions were the visible part, not the problem. Nobody had ever separated delivery unit economics by channel, so the three virtual brands cross-subsidized each other inside a consolidated P&L that landed 45 days late and could no longer drive a single decision. Once we opened cost per order brand by brand, two of the three lost money on every app dispatch and only the third, the one with the highest ticket, covered the hole.
Own channel vs delivery apps, side by side
| BEFORE (baseline, month 0) | AFTER (month 7) | |
|---|---|---|
| Channel mix (apps / own) | ✕94% apps · 6% own channel | ✓61% apps · 39% own channel |
| Intermediation cost over gross sales | ✕26.8% weighted average | ✓17.1% weighted average |
| Prime Cost (food + labor) | ✕71.4% | ✓63.8% |
| Theoretical vs actual food cost variance | ✕6.9 points (theoretical 28.3% · actual 35.2%) | ✓1.4 points (theoretical 28.1% · actual 29.5%) |
| Labor Cost over sales | ✕36.2% | ✓34.3% |
| Consolidated average ticket | ✕9.40 USD | ✓12.10 USD |
| 90-day repeat rate (identified customers) | ✕8% (no owned database) | ✓31% across 4,100 owned contacts |
| EBITDA over sales | ✕-1.9% | ✓8.6% |
| Consolidation window for the result | ✕— | ✓Held for 3 months (months 5, 6 and 7) |
The opening picture: 94% of volume in someone else's hands
A dark kitchen running three virtual brands out of 96 square meters, with 11 people between kitchen and dispatch, closed 780 thousand USD the year before the intervention, with an average ticket of 9.40 USD and 94% of its orders arriving through Rappi, Uber Eats and DiDi Food. The remaining 6% split between WhatsApp and an ordering page built two years earlier that nobody ever promoted. That split is not unusual in the region: Rappi reported 35 million active users and 150 million downloads as of August 2024 (Rappi, operating report, 2024), so the commercial gravity of the platform is real and no amount of willpower cancels it. Low revenue was never the problem. The problem was that nobody had ever separated cost per order brand by brand, and with a consolidated P&L arriving 45 days late, deciding meant guessing late.
Two brands lost money on every dispatch and the third one covered it
Once we opened unit economics by channel and by virtual brand, the cross subsidy hidden inside the consolidated statement showed up: two of the three brands lost money on every app dispatch and only the highest-ticket one held the result together. With intermediation running from 22% to 30% on a 9.40 USD ticket, the platform took between 2.07 and 2.82 USD before food cost, packaging or dispatch labor were even touched. That is why entire cities legislated the matter: San Francisco capped commissions at 15% (Restaurant Dive, 2020) and New York made permanent a maximum of 15% per delivery plus 5% for other services (Restaurant Business, 2023). Where no cap exists, you set your own cap through channel mix. Commission is not an administrative expense, it is a MIX variable.
Buying discovery is not the same as buying repeat purchase
Framing own channel against delivery apps as an either-or choice is the original mistake, and whoever frames it that way lost the margin before negotiating it. The own channel buys something different, REPEAT PURCHASE, and its acquisition cost falls with every repeated order because the customer returns through a door they already know. The app charges its percentage on order one and on order forty, with no loyalty discount; geotargeted advertising and Local SEO cost a lot the first time and almost nothing the tenth. Two different purchases, two different curves.
What we did with the Masterestaurant method, in order?
The intervention started with the Masterestaurant method's unit economics board by channel, the tool that forces you to cost every single order with commission, packaging and dispatch minutes included before offering any opinion about it.
With that costing on the table, Diego F. Parra decided the opposite of what the owner was asking for: do not leave the apps. We raised menu prices on platform only, between 12% and 18%, across the two loss-making brands, kept the real price on the own channel, and slipped into every package a physical coupon worth 15% off, valid exclusively on the ordering page. Geotargeted advertising stayed inside a three-kilometer radius around the kitchen with a monthly budget of 420 USD. None of this is sophisticated. It is arithmetic applied with discipline for six straight months, which is the part almost nobody sustains.
The six-month result: less app volume, more cash
Six months in, the mix moved from 94% platform to 71% platform and 29% own channel, according to the internal case close, with annual revenue projected at 812 thousand USD and a direct-channel average ticket of 12.60 USD against the 9.40 USD weighted ticket at the start. Consolidated contribution margin rose 6.3 percentage points, per that same close. The interesting part sits in the uncomfortable detail: absolute app order volume dropped 9% after the price increase, and the result still improved, because every lost order from the loss-making brands handed money back. Losing orders that cost you money is not a loss. And 40% of own-channel volume came from customers who had ordered through Rappi first, meaning the platform paid for the introduction.
Owning the customer beats a point of commission
The structural difference between the two channels is not commission, it is OWNERSHIP of the contact. On the platform you rent a relationship that ends the day the ranking algorithm changes or the service fee goes up, and neither decision passes through your board. On your own channel you own the phone number, the order history and the context in which the customer buys again. That asymmetry explains why two kitchens with identical food cost end up with opposite EBITDA. The context is hardly marginal either: nearly 75% of restaurant traffic in the United States is already off-premise according to the National Restaurant Association (2025), and more than 20,000 ghost kitchen locations were operating in the United States in 2023 (Statista). At that competitive density, your own customer list is the only asset nobody can reassign.
Transferable lessons by annual revenue band
Under 500 thousand USD a year: this week, calculate cost per order for your best-selling dish on each app, with commission and packaging inside, and if it comes out negative raise the price ON PLATFORM ONLY come Monday. From 500 thousand to 1 million: split the P&L by channel before hiring anyone else, because cross subsidy between brands or between channels is where the profit hides. From 1 to 5 million: build your own ordering page with integrated payment and budget geotargeted advertising within three kilometers; the packaging coupon is the cheapest bridge available. Above 5 million: negotiate volume rates with each platform and hire someone dedicated to the own channel alone. Past 10 million, the celebrity-chef archetype with a large-format venue plays another game, because personal brand generates its own discovery and the app becomes contracted logistics, not the customer's owner.
Limits of this case
I would not expect this result in three contexts, and they are worth stating before somebody copies the playbook without reading the fine print. First, a restaurant with a dining room and a settled neighborhood clientele: the own channel already exists there and it is called the front door, so the upside from digitizing repeat purchase is far smaller than in a hidden kitchen with no facade. Second, in markets with a legal commission cap, where the platform already charges 15% as in New York (Restaurant Business, 2023), differentiated pricing yields considerably less because the spread financing it shrinks. Third, at high tickets: with an average near 24 USD per online order as Ken Research reports for Spain (2025), the relative weight of commission on margin shifts and the arithmetic has to be redone from scratch. Redo it with your own numbers.
What actually separates the two models?
The structural difference is not the commission, it is OWNERSHIP of the customer. On a platform you rent a relationship that ends the day the app reshuffles its ranking algorithm or raises its service fee;
on your own channel you own the contact and the context in which they buy again. That asymmetry explains why two operations with identical food cost can post opposite EBITDA. Acquisition cost behaves in reverse across the two channels. The app charges its percentage on order number one and on order number forty, with no loyalty discount, while geotargeted advertising and Local SEO cost heavily on the first purchase and progressively less afterwards, because a customer who ordered direct returns direct.
What actually separates the two models — in practice?
Demand for third-party app ordering is sitting right there, and that same customer paying full commission month after month is exactly the account worth rescuing.
Regulatory caps changed the ground in some markets and left it untouched in others. New York set a permanent cap of 15% per delivery plus 5% for other services (Restaurant Business, 2023) and San Francisco had limited commissions to 15% back in 2020 (Restaurant Dive, 2020), whereas Latin America has no legal ceiling and the effective rates I read in settlement statements run between 22% and 30% of the ticket. A business model built on a fee your supplier can move unilaterally is not a model, it is a bet. The own channel drags costs the apps hide and rookie operators ignore: owned or subcontracted last-mile logistics, customer support when something goes wrong, the payment gateway, website maintenance.
What actually separates the two models — key points?
Together they weighed 9.3 points of the direct ticket in this case. Still less than half of the 26.8% intermediation was costing, but anyone who plans an own channel at zero cost gets a nasty surprise in month three.
An honest comparison of dark kitchen vs physical restaurant forces one weakness into the open: with no dining room, no storefront and no foot traffic, a dark kitchen has no free discovery mechanism beyond whatever it builds digitally. Statista counted more than 20,000 ghost kitchen locations operating in the United States in 2023, most of them fighting for the same screen space inside three or four applications. At that point Local SEO and the Maps listing stop being marketing and become infrastructure.
Own channel vs delivery apps, criterion by criterion
The mistake: treating the apps as the business
- Success was measured by total monthly orders, with no contribution margin split by channel or by virtual brand.
- Every promotion the platform suggested got accepted (2-for-1, restaurant-funded free delivery) without ever calculating where the discount ate the dish margin.
- Identical menu and identical prices on apps and own channel, so ordering direct gave neither the customer nor the restaurant any advantage.
- The Google Business Profile listing carried no real kitchen hours, no dish photography since 2024 and no own ordering button configured.
- Zero database: four years of operation without a single phone number, email or customer to reach out to on a slow Tuesday.
- The consolidated P&L arrived 45 days after close, when the decision that would have changed the outcome was already six weeks stale.
The right method: apps to discover, own channel to retain
- A P&L by channel AND by virtual brand, showing contribution margin after commission, packaging and dispatch cost, closed weekly.
- Differentiated pricing: the app menu absorbs intermediation cost inside the price, while the own channel offers the same dish with a benefit the customer grasps in three seconds.
- Google Business Profile worked as a free discovery asset, with the order button pointing at the own channel and reviews managed by protocol.
- Systematic contact capture on every direct order, with explicit permission, so demand can be activated on a Tuesday at 4 p.m.
- Geotargeted advertising limited to the real delivery radius and measured against app intermediation cost inside that same radius: if advertising acquires cheaper, advertising wins.
- The apps stay and get optimized, because they remain the best first-purchase engine on the market today; what changes is the role assigned to them.
The numbers from this case, and what they are measured against
“I was convinced my problem was the commission, and for two years I negotiated fractions of a point with platform account managers without moving the result at all. The first weekly P&L by brand showed me that one of my three virtual brands lost 0.80 USD on every app order and had been doing it for fourteen months: 11,000 USD I was pulling out of my own pocket, order by order, congratulating myself because volume kept growing. We shut that brand down in week three, raised platform menu prices and built the own channel properly. Seven months later I sell roughly the same units and I keep 8.6 points of EBITDA where there used to be a hole.”
Composite case for illustration: the names and figures in it do not describe a real business and are not industry data.
The intervention timeline, frictions included
We rebuilt the business model brand by brand with the Restaurant Model Canvas and forced one uncomfortable question: what is left after commission, packaging, dispatch cost and funded promotion on the average order of each channel. The baseline came back raw. Weighted intermediation cost of 26.8%, actual food cost of 35.2% against a theoretical 28.3% —a 6.9-point variance nobody was watching— and Labor Cost at 36.2%, for a consolidated Prime Cost of 71.4% when the healthy range for a dark kitchen in this band sits between 60% and 65%. The friction here was data, not method: settlement statements from the three platforms arrived in different formats and two of them did not break out restaurant-funded discounts, so fourteen weeks had to be rebuilt by hand before a single figure could be trusted.
We shut down the virtual brand losing 0.80 USD per order and rebuilt the platform menu with a margin that absorbs commission inside the price rather than outside it. This is where almost everyone gets it wrong: raising app prices is not gouging the customer, it is refusing to subsidize an intermediary with your kitchen's margin. In parallel we opened a real, visible advantage on the own channel —an add-on with high perceived value and low cost— so ordering direct carried a reason the customer understood without reading a paragraph. Friction showed up fast. Prices went up, ranking on one of the apps fell and volume dropped 19% in ten days. We held for three more weeks, compensated with geotargeted advertising into the own channel, and total volume recovered by month two.
A dark kitchen without a storefront has to manufacture its discovery in Maps and in search, or it stays at the mercy of platform algorithms. We reworked the Google Business Profile listings of the two surviving brands with real kitchen hours, fresh photography of twenty dishes, a corrected primary category, delivery attributes and an order button pointing to the own channel; alongside that, a review protocol lifted the rating from 4.1 to 4.6 stars in eleven weeks. Geotargeted advertising was confined to the REAL delivery radius of 4.2 kilometers and judged by a single rule: cost per acquired order had to land below equivalent app intermediation cost inside that same radius. Whenever it failed the test, we killed the ad set without debating it.
The own channel stopped being a forgotten website and started operating like a business: contact capture with explicit permission on every order, repeat tracking at 30, 60 and 90 days, and demand activation in weak dayparts through the Demand Radar, which crosses order history, weather and the local calendar to say which Tuesday deserves a push and with which product. One house rule belongs here, because the question always comes up: if the operation has a dining room, the QR menu is a COMPLEMENT and the physical menu stays, because the printed menu governs service pace and suggestive selling while the QR handles delivery, accessibility and price updates. In a pure dark kitchen the QR lives on the packaging as the door into the own channel, and it worked: 1,900 of the 4,100 contacts came in that way.
We installed a thirteen-line weekly close —sales by channel, effective commission, actual versus theoretical food cost, Labor Cost, contribution margin by brand— that the owner reviews every Monday in forty minutes with his head chef. That ritual is what holds the result once the consultant leaves, and it is also the hardest thing to install, because an operator used to reading a P&L at 45 days experiences a weekly figure as an accusation before experiencing it as a tool. The theoretical-to-actual variance fell from 6.9 to 1.4 points, Prime Cost closed at 63.8% and EBITDA held above 8% through months five, six and seven, which is the minimum window I demand before calling any result consolidated.
And with AI?
Optimize channels, pricing and unit economics of your dark kitchen. Diego F. Parra is an expert in AI applied to restaurants.
Free tools: own channel vs delivery apps
The Masterestaurant tools that carried this case
None of this ran on improvised spreadsheets or on a bespoke dashboard that dies the day the consultant walks out: we used closed, off-the-shelf products, the same ones any operator can deploy next week. The selection criterion was blunt: a tool had to produce a figure that changed a decision within seven days, or it did not make the cut.
Diego F. Parra — International consultant, expert in creating and scaling restaurants and in AI applied to restaurants, foodtech and HORECA. Methodology applied in 8.400+ restaurants across 43 countries · Expert in Artificial Intelligence applied to restaurants, hospitality and food businesses · 20+ years in restaurants, catering, large events and business growth · Author of 3 ISBN-registered books: «Triunfar o morir en el intento» (2013) and «De esclavo a dueño» (2023) · International keynote speaker for the HORECA sector.
Questions I always get about this case
Should I quit Rappi and the delivery apps and sell only through my own channel?
Should I quit Rappi and the delivery apps and sell only through my own channel?
No, and quitting cold is the most expensive mistake I see in dark kitchen from scratch projects. Apps are today's best first-purchase engine: Rappi reported 35 million active users in 2024, and you will not buy that audience with advertising in six months. The right play is keeping them at a price that absorbs commission while migrating REPEAT purchase to your own channel.
What does selling on Rappi, iFood or Uber Eats really cost?
What does selling on Rappi, iFood or Uber Eats really cost?
Nominal commission is only part of it. Add restaurant-funded promotional discounts, reinforced packaging and the cost of orders with incidents, and real intermediation cost lands between 22% and 30% of the ticket in Latin America. New York caps it at 15% per delivery by law (Restaurant Business, 2023); the region has no ceiling at all.
How long does an own channel take to become profitable against the apps?
How long does an own channel take to become profitable against the apps?
In this case the own channel passed break-even in month four and reached 39% of volume by month seven. The timeline depends on three variables: population density inside the delivery radius, quality of the Google Business Profile listing, and discipline in capturing contacts. Without an owned database there is no own channel, only a website.
Does this method work for a physical restaurant or only for a dark kitchen?
Does this method work for a physical restaurant or only for a dark kitchen?
It works for both, with an advantage for the physical venue: it already has a storefront, foot traffic and a dining room, which are three free discovery mechanisms a dark kitchen must build from zero. If you have a dining room, keep the physical menu to govern service and use the QR as a complement for delivery and price updates, never as a replacement.
2026 data on own channel vs delivery apps
Verifiable industry benchmarks from official, non-commercial sources (government, industry associations, market research) - not competitors.
| Metric | Value | Source |
|---|---|---|
| UAE cloud kitchen market | US$ 430 millones (2025), proyectado a US$ 1.082,6 millones en 2032 (CAGR 14,1%) | Coherent Market Insights 2025 |
| DoorDash US delivery market share | 60.7% of the market at the end of 2024 | Earnest Analytics 2024 |
| Uber Eats US delivery market share | 26.1% of the market at the end of 2024 | Earnest Analytics 2024 |
| Grubhub US delivery market share | 6.3% of the market at the end of 2024 | Earnest Analytics 2024 |
| Glovo q-commerce turnover | More than €1,000 million a year, with retail and grocery growing ≈50% in 2024 | EU-Startups 2025 |
| Blinkit dark stores in India | ≈2,100 dark stores, with plans to add 900 more by March 2027 | Storyboard18 2025 |
Related content
Own channel vs delivery apps: repeat this case in your restaurant
Applied in +8.400 restaurants across 43 countries.