A ghost kitchen is a kitchen that cooks only for delivery and pick-up, with no dining room; customers find it on apps. An existing restaurant can test one without a new site by listing a small second menu, cooked in its current kitchen, on one delivery app for four weeks, capped by capacity and judged on profit per order.
Dark kitchen, cloud kitchen and virtual restaurant are other names for the same delivery-only idea. It comes in three set-ups, and only one of them needs no new premises, which is the one tested here: a virtual brand, meaning a second, delivery-only menu run from a kitchen you already have.
One criticism deserves attention before you start. Some virtual brands have been criticised for selling food from an existing restaurant under a different name without being open about it. The test below is built on being plain about who cooks the food, and on protecting the customers in your dining room, because a test that damages the main business has failed whatever its orders say.
Ghost, dark, cloud, virtual: one idea, three set-ups
The names are used loosely, so it helps to separate the set-ups by what you have to build or rent.
| Set-up | What it is | What you commit to | Fits a four-week test? |
|---|---|---|---|
| Shared facility | Several brands cook in one rented kitchen with no dining rooms | A new site, a rental agreement and your own staff or a partner's | No: the commitment comes first |
| Standalone delivery-only kitchen | One brand in a separate kitchen set up only for delivery | A lease, equipment, licences and a team | No |
| Virtual brand in your own kitchen | A second, delivery-only menu cooked by your current team and equipment | A menu, photos, packaging and some staff attention | Yes |
The third is the only one you can switch off in a day. If it works and your kitchen becomes the bottleneck, the next question is whether a separate kitchen would pay, and the cash-flow method in are ghost kitchens still worth it is the tutorial for that. This one stops at the point where you know whether the idea earns money in the kitchen you already have.
"Without building a new site" also means without a new website. The delivery app is the storefront, with its own menu, photos and ratings, so there is nothing to design, host or promote outside it.
What one delivery order really earns
Delivery-only economics turn on the commission. As of October 2026, DoorDash's merchant pricing lists delivery commissions of 15%, 25% or 30% of the order subtotal depending on the plan (Basic, Plus or Premier), and 6% on pick-up. Those are its USD list rates; other apps and other markets differ, so put the rates from your own agreement into the same sheet. The dearer plans are sold on wider reach, so check in the merchant portal what each one adds before assuming the top rate pays for itself. Commission is charged on the subtotal after any discount you fund, and promotions and ads you opt into carry their own fees.
The table follows one order with a $26.00 subtotal through the costs. Every figure except the commission rates is illustrative: food cost at 31% ($8.06), packaging $1.10, extra labour $2.00 and an allowance for remakes and refunds of 3% ($0.78). Replace each with your own.
| Route | Commission | Left after food, packaging, labour and remakes | Share of the subtotal |
|---|---|---|---|
| Delivery on Basic (15%) | $3.90 | $10.16 | 39.1% |
| Delivery on Plus (25%) | $6.50 | $7.56 | 29.1% |
| Delivery on Premier (30%) | $7.80 | $6.26 | 24.1% |
| Pick-up (6%) | $1.56 | $12.50 | 48.1% |
Two sums are worth doing on your own numbers. First, the promotion effect. On the Plus route, a $5 discount that you fund cuts the subtotal to $21.00 and the commission to $5.25, but the food, packaging and labour cost the same, so what is left falls from $7.56 to $3.81, a drop of about half. Second, the extra-labour line. If your cooks are on the clock anyway in a quiet period, the true marginal cost of an order may be close to zero. Put a modest figure in all the same, so the test does not flatter itself, and ask afterwards whether the extra orders forced any extra hours.
The tutorials on cloud kitchen margins per brand and on whether takeaway menu prices should rise on apps go further on both points. For the test, the figure that matters is the third column: what each order leaves after everything it costs.
The no-new-site test: one app, one menu, one quiet daypart
A test is only low-risk if its limits are set in advance. These are the choices to make before listing anything.
| Element | Choice for the test | Reason |
|---|---|---|
| App | The one your own customers already use for delivery | One app keeps the data clean and saves set-up time |
| Menu | Six to eight items, all from ingredients you already buy, with no more than two new ones | Small menus cook fast and travel well, and your purchasing barely changes |
| Daypart | The quietest part of your week, such as weekday mid-afternoon | Idle kitchen capacity is the only free resource you have |
| Capacity cap | A fixed number of tickets per 15 minutes, and a rule to pause the listing when the dining room is busy | The dining room comes first |
| Price | Set so the contribution works on the plan you actually pay, not on pick-up | See the table above |
| Length | Four weeks, with a review after each | Orders on a new listing build slowly, so one week tells you nothing |
| Budget | Set-up and packaging stock written down in advance | You need a figure to recover |
On the app question, ask the app's support how to add a second store at the same address under your existing account, and whether it needs a separate onboarding review or a registered trading name. Also ask whether any introductory offer reduces commission in the first weeks, because a lower rate flatters the results and should be recorded in the sheet.
Pricing a virtual brand is slightly freer than pricing your main menu, because there is no dining-room price to match, and slightly riskier, because a delivery shopper sees your items beside a dozen others in the same app. Look at what similar items cost there, since your customers will, and then set your own prices from your own costs and the contribution sums above. Do not set a price you could not defend from your invoices.
Choose the menu from what the kitchen already stocks
An AI assistant is useful here, because overlap between ingredients is a matching job. Give it your real ingredient list, equipment and limits. Take, as an example, a 40-seat bistro with a fryer, a flat-top, an oven and a cold station, quiet on weekdays between 15:00 and 17:30.
I run a 40-seat bistro. Ingredients we already buy every week: chicken thighs, brioche buns, potatoes, cheddar, pickles, mayonnaise, lettuce, tomatoes, onions, eggs, flour, breadcrumbs, buttermilk. Equipment: fryer, flat-top, oven, cold prep station. We want a delivery-only menu cooked on weekdays from 15:00 to 17:30.
Constraints: no new equipment; no more than two new ingredients; at most eight items; each item cooks in under 10 minutes; each item must still be good after 20 minutes in a bag. Suggest one concept and eight items. For each item list the ingredients, the new ones, the travel risk, and the allergens that appear in the ingredient list only. Do not estimate costs, and do not make any dietary or health claim.Suppose the reply proposes a single concept, here crispy chicken, and a list of items you can check against your own costs. The costs below come from the bistro's invoices, not from the assistant.
| Item | New ingredient? | Travel risk | Cost | Price | Food cost |
|---|---|---|---|---|---|
| Crispy chicken sandwich | None | Low | $2.95 | $9.50 | 31% |
| Cheddar crispy chicken sandwich | None | Low | $3.35 | $10.50 | 32% |
| Chicken tenders, three | None | Low | $2.40 | $8.00 | 30% |
| Tender box with fries | None | Medium | $3.30 | $11.50 | 29% |
| Fries | None | Medium | $0.85 | $4.00 | 21% |
| Slaw | Cabbage | Low | $0.70 | $3.50 | 20% |
| Hot honey dip | Hot honey | Low | $0.35 | $1.00 | 35% |
Three things to fix in a draft like that:
- An assistant may offer a "gluten-free option" for the fries. They share a fryer with breaded chicken, so no such claim can be made. Delete it. Allergen information comes from your ingredient sheet and your cross-contact practice, never from the draft.
- If you let it estimate costs, it uses generic market prices. Replace every figure with your invoice price and recalculate. The tutorial on supplier invoice price creep explains why last quarter's invoice is a poor guide to this quarter's.
- If a draft or a staff member labels the slaw "vegan" because cabbage sounds plant-based, remember that the dressing may be mayonnaise, which contains egg. Check each label claim against the ingredient list. The hot honey dip also fails the 33% rule of thumb you may set for food cost, so reprice it or drop it.
Names need the same care. Suggested brand names are often already in use, so search the app and your business register before committing to one. The best test menu is boring on purpose: if every item uses what the kitchen already does well, you are testing demand, not your own ability to learn new recipes.
Three other kitchens, three other menus
The bistro's chicken menu is one answer. The method fits any kitchen with an idle period, but the binding constraint changes from kitchen to kitchen, and finding it is half of the design.
- A pizzeria with one deck oven. The oven is the bottleneck, so the cap is set in oven slots rather than tickets, and the test watches pizza ticket times first. A small "wings and garlic bread" brand that bakes at the same temperature as the pizzas is easier to fit than anything needing a fryer the kitchen does not have.
- A café that closes at 15:00. It has grill and toaster capacity left in the afternoon, so a soup-and-toastie brand for late afternoon is natural. The risk is waste, because soup and fillings made for a slow afternoon end up in the bin. Start with two soups in small batches and one toastie format, and count what is thrown away as a cost of the test.
- A grill restaurant that is busy at weekends and quiet midweek. A delivery-only burger brand could run on weeknights only and switch off from Friday to Sunday. The weeknight-only rule protects the weekend service that earns the most.
In each case the checklist is the same: one app, six to eight items, existing ingredients, a cap, a pause rule and four weeks of numbers. What changes is the unit the cap is counted in, whether tickets, oven slots or litres of soup.
Honest listings: real photos, true names, allergens from the sheet
The listing is the whole storefront, and three rules keep it honest. First, photos show the real item. DoorDash's photo guidance asks sellers to use real photos of the actual item, and an AI-generated image of a dish you do not make that way is a likely route to complaints and refunds. A phone photo in daylight on a plain plate does the job. The tutorials on AI food photography on menus and delivery apps and on rewriting delivery app menus go further.
Second, descriptions say what the item is. AI is good at tidying a draft, and it is a poor source of facts about your food. A filled-in listing might read:
[Brand name]: crispy chicken, cooked to order.
Made in the kitchen of [restaurant name].
Crispy chicken sandwich: buttermilk-marinated chicken thigh, pickles, lettuce and mayonnaise on a brioche bun.
Contains: gluten (wheat), egg, milk.
Cooked in a kitchen that also handles other allergens. Ask us before ordering if you have an allergy.Third, the "Made in the kitchen of" line is my suggestion, and it answers the transparency criticism directly. Check that the app's format allows it, and do not invent a backstory such as a family recipe or a founder you do not have. The address the app shows will be your real one, so a story that contradicts it only invites distrust.
Protect the dining room: capacity, safety and paperwork
Choose a rule for pausing before the first order arrives. Look for the app's pause and prep-time controls in your merchant portal, and decide that the listing goes off whenever the dining room has more than a set number of tables, or the open tickets reach your cap. Then measure the dining room itself: average ticket time before the test and during it. If it slips by more than a margin you set in advance (two minutes is used in the example below), pause and review.
Food safety does not change because the customer is not in the room. The same cooking, holding and cross-contact controls apply, and allergen information must be as accurate as on your main menu. Paper matters too. Check with your licensing office whether a second trading name or delivery-only sales need registering, and tell your insurer that you are adding delivery. Rules differ by place and the answer can take a week, so ask before the first listing.
Then plan the packing, because most of what goes wrong in delivery happens between the pass and the driver's bag. Name one person to run the delivery tickets in the quiet daypart, with a pack-check routine: read the ticket aloud, count the items, add the sauces and cutlery, check the lid and seal the bag, and stamp the time. Keep a refund log from day one with the date, the order, and the reason (missing item, cold food, wrong item, spill or taste). A week of entries tells you whether the fix is the menu, the packaging or the routine.
The bistro's four-week log, week by week
The bistro lists its menu on one app, on the Plus plan at 25%, on weekdays from 15:00 to 17:30. Set-up costs are $120 of packaging stock and eight hours of staff time at $16, which comes to $248. Each order leaves 41% of its subtotal after commission (25%), food cost (31%) and remakes (3%), less $3.10 for packaging and extra labour. In a sheet: =0.41*B2-3.10, with the average subtotal in B2.
| Week | Orders | Average subtotal | Left per order | Left in the week | Running total |
|---|---|---|---|---|---|
| 1 | 9 | $22.40 | $6.08 | $54.76 | $54.76 |
| 2 | 13 | $23.60 | $6.58 | $85.49 | $140.24 |
| 3 | 18 | $24.90 | $7.11 | $127.96 | $268.21 |
| 4 | 21 | $25.30 | $7.27 | $152.73 | $420.94 |
Read it in order. Across 61 orders the test left $420.94, an average of $6.90 an order. The set-up cost of $248 was recovered during week 3, when the running total passed it at about the 38th order. After set-up costs the four weeks are worth about $173. That is a modest result, and a real one only if three other checks hold.
- The dining room. Average dine-in ticket time was 14.0 minutes before the test and 14.6 during it, inside the two-minute margin. The listing was paused twice on busy evenings.
- The main menu. The restaurant's own delivery orders on the same app ran at 62 a week before the test and 61, 60, 63 and 59 during it. With a count of that size, chance alone moves it by around eight a week, so there is no sign that the new brand stole orders the main one would have had.
- Quality. Two of the 61 orders were remade or refunded (3.3%), against the 3% allowed in the sums, and the first 12 ratings averaged 4.5.
Where AI helps in the test, and where it must stay out
The assistant is an analyst and a drafter in this test, never the person who decides or the source of a fact about your food.
- Reading the order data. Download the app's order export if your merchant portal offers one, remove customer names and addresses, and ask for weekly orders, average subtotal, items sold by count and refund rate, with formulas shown. Check the totals against the portal's own summary, because an assistant can mis-sort a date column. It cannot see your dining room, so supply the ticket times and the main-menu counts yourself.
- Summarising comments. Paste the written ratings with names removed and ask it to group them into themes, count them and quote two examples each, without inventing any.
- Checking the costing sheet. Ask it to recalculate each food-cost percentage and flag any item above your threshold, with the working.
- Tidying descriptions. Let it improve phrasing you have already checked against the ingredient sheet.
An illustrative result of the comment task, on 14 written comments from the first two weeks: cold on arrival 4, missing item 3, portion felt small 2, taste praised 5. The numbers add up to 14, and the pattern points at the fix: a shorter gap between the pass and the driver, and the pack-check routine, before anything about the recipes. Read the originals yourself as well, since a summary can flatten the one comment that matters.
Keep the assistant away from four things. Allergen statements and dietary or health claims come from your ingredient sheet and your kitchen practice, not from a draft. Supplier prices come from invoices. Photographs must show the dish you actually make. And the go, change or stop decision stays with a person who has read the numbers.
Go, change or stop: the rules to set before week one
Write the decision rules down before the first order, while you are still neutral. These are the bistro's thresholds; yours should come from your own sums.
| Check | Go on (extend or launch) | Change something | Stop |
|---|---|---|---|
| Left per order | $6.00 or more | $3.00 to $6.00: raise prices, cut slow items or trim promotions | Under $3.00 after changes |
| Orders in week 4 | 20 or more | 10 to 19: sharpen the menu and photos for four more weeks | Under 10 |
| Dine-in ticket time | Up less than 2 minutes | Up 2 to 4 minutes: tighten the capacity cap | Up more than 4 minutes |
| Remakes and refunds | 5% or less | 5% to 8%: fix the weak items or packaging | Above 8% |
| Main-menu orders | Within normal variation | Falling for three weeks running: investigate overlap | A clear fall linked to the test |
Against that table every check on the bistro's result says go on: $6.90 left per order, 21 orders in week 4, ticket time up 0.6 minutes, refunds at 3.3% and the main menu flat. "Go on" still does not mean a full launch, because 21 orders a week is a promising start and not yet a business. The bistro's next step is another four capped weeks with two changes, dropping the slowest items and adding a combo, and only then a decision on growing the brand.
Whatever you find, write the numbers down, tell your team what happened and why, and keep the listing's economics in the same sheet as the rest of the restaurant. If you would like help setting the thresholds from your own figures, that is what my AI implementation consultation covers, and the wider cash-flow question of whether a delivery-only business is still worth it lives in the tutorial linked above.
Ghost-kitchen questions restaurant owners ask
Is a ghost kitchen the same as a cloud kitchen?
In everyday use, yes: ghost kitchen, dark kitchen, cloud kitchen and virtual restaurant all describe a delivery-only food business. 'Virtual brand' usually means a delivery-only menu run from a kitchen that already serves another brand, and 'shared kitchen' means several brands cooking in one rented facility.
Do I need a new licence or trading name to test one?
It depends on where you trade. A second trading name, delivery-only sales or a different menu can change what your registration, insurance and food-safety plan must cover. Ask your licensing office and your insurer before listing anything, and keep allergen information as accurate as it is on your main menu.
Which delivery app should I test on?
The one your own customers already use for delivery. A single app keeps the data clean, saves set-up time and lets you compare the test with your existing delivery orders. Add a second app only after the test, so you can tell which one earned what.
Can AI write the menu and make the photos?
AI can suggest items that fit your ingredients, draft names and descriptions and check the costing sheet. It must not write allergen statements or health claims, and food photos should show your real dishes: DoorDash's photo guidance asks for real photos of your actual item. Review every line before it goes live.
Further reads
- Cloud Kitchen Inventory: Forecast Prep by Delivery Platform With AI — Forecasting prep by delivery platform once a brand is running.
- AI for a Cloud Kitchen: Keep Every Brand's Menu Consistent on Apps — Keeping menus consistent across several apps.
- AI Ordering vs Delivery Apps: What a Takeaway Actually Pays — What a takeaway actually pays on apps versus its own ordering.
- Restaurant Menu Price Increase: Test Each Change by Dish With AI — Testing menu price changes dish by dish.
- AI Chatbot for Restaurants: Hours, Bookings and Hand-Offs to Staff — Answering the dining room's own questions, with hand-offs to staff.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Wikipedia's article on virtual restaurants (definition, set-ups and criticisms); DoorDash's merchant pricing, promotions and photo-guideline pages (USD plan rates and discount treatment, checked October 2026); Toast and DoorDash descriptions of virtual brands. The bistro and every worked figure are illustrations, so replace them with your own costs.