Takeaway Delivery App Menus: Rewrite Descriptions and Photos With AI

Give AI a fixed formula for every description (contents, character, count, catch), check each line against the kitchen, then pair the words with real photos that match what you send out. Use AI to tidy light and crop, never to invent a dish. Change ten items, leave ten alone, and compare two weeks of orders.

The words are the quick part. A model has never tasted your food, so left alone it will write "freshly made" over a bought-in base, or "mild" over a dish that sends people reaching for the water jug, and on a delivery app that gap arrives as a refund request rather than a shrug. Use AI for the drafting, and leave the final say on every claim with the kitchen.

Follow me on Instagram@sagnikteaches

Pick ten items to fix first, not the whole menu

A 90-item menu rewritten in one go cannot be tested, and mistakes hide in the volume. Start with a short list drawn from the last 28 days of sales: the item report in each app's merchant portal where it offers one, or an export from your till. Look for three kinds of item.

Connect on LinkedInSagnik Bhattacharya

• Best sellers with weak photos. A small lift on a dish that already sells a lot is worth more than a big lift on a dish nobody orders.
• Descriptions that only repeat the name. "Chicken wings: chicken wings" tells a customer nothing about heat, sauce or how many arrive.
• Anything that attracts "not as pictured", "wrong sauce" or "missing item" complaints. Those are photo and option-list faults, and each one costs more than a lost order does.

Put what you find into a quick audit before you rewrite anything. This is the version I would start with (illustrative; add signals of your own).

Subscribe on YouTube@sagnikteaches

Which dishes deserve a push is a margin question, and AI menu engineering: which dishes to promote, reprice or drop covers it. Here the filter is simpler: sells already, could sell better. Allow an afternoon for ten items: about 45 minutes to export and audit, an hour to draft and check with the chef, 45 minutes for photos and 30 minutes to enter the changes in each app.

A four-part formula for descriptions that work on a phone screen

Good delivery copy answers four questions in the order a hungry person asks them: what is in the box, what does it taste like, how much is there, and is there a catch. I call the parts contents, character, count and catch.

• Contents: the physical things in the box, named plainly and taken from the recipe card rather than from the AI's imagination.
• Character: one taste or texture word that separates this dish from its neighbours on the menu, such as crisp, smoky, creamy, sharp or mild.
• Count: pieces, inches, grams if you weigh portions, or "serves two".
• Catch: the heat level, what is not included, how it travels or how long it takes at busy times. Leave it out when there is nothing to say.

Aim for 12 to 25 words and put the contents in the first line, because a phone screen can cut a long description off after a couple of lines (look at your own menu on a real phone to see where). Customers scan and search for plain words such as wings, curry or vegetarian, so keep the plain word in the item name and let the description carry the detail. If your counter menu runs on numbers ("No. 23"), lead the app name with the dish, because an app customer has never seen your number board. Check the menu editor for a character counter and hand that limit to the AI.

Here is the formula applied to five dishes from an imaginary mixed-menu takeaway. The wording is illustrative, and every real line needs the chef's tick.

The "after" column has no "delicious", no "authentic" and no "famous". Those words carry no information, and "authentic" or "homemade" are claims a customer can hold you to. For the wider question of keeping a menu in your own voice, see how to write menu descriptions with AI that still sound like you.

Prompts to paste in, a sample result and what needs fixing

ChatGPT, Claude and Gemini can all do this on a free plan. The quality comes from what you feed them, so give the model your recipe card lines and it has nothing to invent. Work ten items at a time, as a plain list.

You are writing item descriptions for a takeaway's delivery-app menu. Customers read them on a phone and decide in a few seconds.

Use this order for every description: contents, character, count, catch.
- Contents: what is physically in the box, taken ONLY from the recipe line I give you.
- Character: one taste or texture word that separates this item from similar ones.
- Count: size, number of pieces, or who it serves.
- Catch: heat level, what is not included, or how it travels. Leave it out if there is nothing to say.

Rules:
- 12 to 25 words each, no exclamation marks.
- Use the words my customers use: [chips or fries], [aubergine or eggplant].
- Never use: delicious, mouth-watering, authentic, famous, best, homemade, fresh, healthy.
- Do not add any ingredient, claim, dietary label or allergen statement that is not on the recipe line.
- If something you need is missing, write ASK KITCHEN: and the question, instead of guessing.
- Start each description with a different word from the one before.

Items (name | recipe line | size | heat):
1. Chicken katsu curry | breaded chicken breast, steamed rice, mild curry sauce, pickled ginger | single portion | mild
2. Smash burger | two beef patties, cheese, pickles, house sauce, bun | single burger, no sides | none
3. Loaded fries | fries, cheese, spiced beef, jalapeños | [size needed] | medium

A typical first pass looks like this (illustrative output, with my notes under it).

Chicken katsu curry: Crisp breaded chicken breast on steamed rice with a mild, gently sweet curry sauce and pickled ginger. One hearty portion.

Smash burger: Two pressed beef patties, melted cheese, pickles and house sauce in a toasted brioche bun. Juicy and stacked high. Chips not included.

Loaded fries: Fries under melted cheese, spiced beef and sliced jalapeños, with a medium kick. ASK KITCHEN: how big is the portion, and are the fries skin-on?

Three things need fixing. "Brioche" was not on the recipe line, so the model supplied it, and either the chef confirms the bun or the word goes. "Juicy and stacked high" is filler that the photo says better. "Gently sweet" is a taste claim, so ask whoever makes the sauce. The ASK KITCHEN line is the prompt working as designed: the model declined to guess a portion size, which beats a confident wrong answer.

When the chef is happy with the wording, run a second pass that plays to what AI is good at, which is listing what you have promised.

Below are final menu descriptions. List every claim a customer could hold us to (freshness, origin, dietary suitability, heat level, size, timing, "homemade" or similar), one per line, with the item name. Do not judge the claims. Just list them so the chef can tick or strike each one.

The chef can clear that list in a few minutes. Anything labelled vegan, halal, gluten-free or similar belongs on a different list: take dietary and allergen wording from your own allergen records and from the allergen label tutorial, never from a model's phrasing.

Photo rules by app, and one shoot routine that suits them all

Each app moderates photos on its own terms, so the practical approach is one routine for every dish plus a short rules box per app. DoorDash and Uber Eats are the examples below; add any other app you list on in the same way.

The shoot routine for every dish

1. Photograph the dish exactly as you pack it: the same portion, sides and sauces, with no props that do not come in the order. DoorDash rejects photos that do not match the item.
2. Use window light, and switch off the flash and the yellow overhead lights. Wipe the container rim and the table first.
3. Shoot landscape with the dish centred and space around it, then crop to each app's ratio later. Shooting wide leaves you room to crop.
4. Take one photo per different dish. The beef burger's picture should not stand in for the chicken burger.
5. Keep text, logos, price stickers and watermarks out of the frame.
6. Use the same surface and angle across the menu, so the set looks like one shop.
7. Export at full size, then check the file against each app's size cap.
8. View the result on a phone at actual size. If you cannot name the dish within a second, retake it.
9. Name files by item number, such as 014-katsu-curry.jpg, so the right picture lands on the right item.

A shot-list prompt saves a second session. Paste your ten items into the model and ask it to group them by container, angle and background, so one 45-minute session covers the lot.

What each app says about photos

Rules change, so treat this table as a nudge to read the live page, not a substitute for it. The DoorDash details come from its help pages as of October 2026, and those pages do not all agree with each other, which is one more reason to check inside your own portal.

DoorDash lists its common rejection reasons on its own help page. Photos that do not match the item, duplicates, text or logo overlays, blurry shots, crops so tight that part of the item is cut off, and distracting or artificial-looking backgrounds are among them, which is worth knowing before you reach for an AI filter. Turned into a pre-upload check for DoorDash, it is six lines you can tape beside the monitor:

1. Landscape 16:9, at least 1400 × 800 pixels, and within the file-size limit your portal shows.
2. Dish centred, the whole item visible after cropping, with enough space around it that a square thumbnail still shows the dish.
3. Shows exactly what comes with the item, and nothing extra.
4. No text, logo, border or watermark added on top.
5. Sharp focus, natural colours and a real, uncluttered surface behind the dish.
6. A separate photo for each different item.

Build the same short list for Uber Eats from its own guidance and keep the two side by side. Where the lists differ, shoot to the stricter one and crop a separate copy for each app.

Using AI on photos without misleading anyone

AI earns its place in the finishing, not in the dish. Anything that changes what the customer will actually receive is the line you stay behind.

A realistic mistake shows why the line matters. Say a burger shop edits its best seller to add a second patty because the single looked small. Orders rise for a week. Then come "not as pictured" reviews and refund requests, which can cost more than the extra orders earned, and a mismatch with the item is one of the reasons DoorDash gives for turning a photo down. The honest fixes are to serve and price the bigger burger, or to shoot the burger you do serve from a better angle.

Background swaps need the most restraint. DoorDash's own editor offers background changes, yet its rejection guidance lists distracting or artificial backgrounds, so pick a plain, believable surface such as a wooden board or a matt grey worktop, never a scene, and let the app's review have the final word. Better still, clear the worktop before you shoot.

You do not need paid software for the finishing. A phone's own editor handles light and crop. DoorDash's help pages describe a photo editor that opens after upload, and an AI Photo Enhance tool with three modes (AI Retouch, AI Replate and Match Style) whose results still go through DoorDash's normal photo review; check whether your portal offers it. Whatever tool you use, compare the finished image with a packed order before you publish, because a replated dish that looks better than the box you send is the same problem as an invented one. Canva's free plan allowed up to 20 AI tool uses a month when its pricing page was checked on 27 October 2026, which covers a ten-item batch. The bigger question of AI-made food images, and what the apps and your customers will accept, is in can a restaurant use AI food photography on menus and delivery apps.

Modifier groups: fix the option screens before the wording

Modifier groups are the option screens that open when a customer taps an item: size, base, sauce, toppings, extras, "make it a meal". They decide whether the kitchen gets a usable ticket, and they are where many menus quietly lose orders or create remakes. The usual faults:

• Five or six screens for one pizza, so ordering drags.
• The same option under two names ("Extra cheese" in one group, "Add cheese" in another), sometimes at two prices.
• A sauce choice marked optional, so the wings arrive plain.
• A group called "Dips" with no maximum, so one box can be ordered with eight free pots.
• Thirty toppings in one unsorted list, with the popular ones at the bottom.

Pattern-spotting across a table is a job AI does well. Copy one item's groups into a sheet (group name, required or not, minimum, maximum, option, price change) and paste it under a prompt like this.

Here are the modifier groups for one takeaway menu item, as a table: group name, required (yes/no), minimum, maximum, option, price change.
1. Flag duplicate or near-duplicate options across groups.
2. Flag any group whose name does not match its minimum and maximum (for example "Choose 2" with a maximum of 5).
3. Flag any group the kitchen needs answered but that is marked optional, and any required group with no sensible default.
4. Propose a tidier structure of no more than four groups, as a table, with group names written as instructions ("Pick your base", "Choose up to 3 dips").
Do not change any price. End with a list of every change you made so I can re-enter them by hand and check each one.

For a 12-inch pizza that began with eight screens (size, base, sauce, cheese, two topping lists, extras and dips), a sensible result is three:

Size became item variants, and sauce and cheese moved into the item itself. Two warnings apply. The model cannot see your app's limits on groups, options or nesting, and they differ by app, so it will sometimes propose a layout the editor refuses. And pricing rules live in your till and your app, not in the sheet, so enter the changes by hand. Then add the item to a basket on a phone and step through every screen as a customer would, on each app.

A two-week test with ten items and ten controls

Rewrites feel like an improvement whether or not they are one. A fortnight with ten rewritten items and ten left alone shows which it was. Two weeks covers two weekends, which evens out busy and quiet days.

1. Pick ten items to change and ten similar items to leave alone, matched on price and popularity.
2. Write down the last two fortnights for both groups: total orders and units sold of each group.
3. Make every change on day one, for the test group only. Photos need approving first (DoorDash's newer help pages say about one business day, and some other pages say longer), so start the count when the new picture shows live.
4. Change nothing else for 14 days: no price moves, no new offers, no reordering of categories. Log what you cannot control, such as a public holiday, a rainy weekend or an app promotion.
5. Compare units sold per 100 orders, not raw units, because order volume moves from week to week.

Here are illustrative figures for a burger-and-wings takeaway taking about 1,100 app orders a fortnight across two apps.

The rewritten items moved from 5.85 to 6.06 units per 100 orders between the two earlier fortnights, so the normal swing is about 0.2. In the test fortnight they reached 7.36, about 1.4 above their old average of 5.96, a lift of roughly 24%. The control items went from an average of 5.77 to 5.82, a lift of about 1%. A gain nearly seven times the usual swing, with the control group sitting still, says the rewrite did the work and not a busy fortnight.

In money, the ten items would have sold about 66 units at the old rate and sold 81, so roughly 15 extra. At an average price of $11.50 that is about $175 of extra sales a fortnight, and the app's commission comes off it before food cost (15%, 25% or 30% of the order subtotal on DoorDash's published plans). That is modest for an afternoon of work, and it is not a number to multiply across the whole menu, because gains differ by dish. That is the reason to test ten at a time.

The decision rule is short. Keep the changes when the test group's lift beats the control group's lift by more than the biggest swing between the earlier fortnights, and complaints about those items have not gone up. If the lift is smaller than the usual swing, you have learned nothing yet, so run another fortnight before deciding. Price moves belong in a separate test, and the delivery app price markup tutorial explains how app prices relate to counter prices.

Seven faults in AI menu copy to catch before a customer does

These are the failures I would expect in a first batch. Each comes with the quickest way to spot it.

• An invented ingredient. The brioche bun above. Tick every noun in the output against the recipe card.
• A freshness or origin claim you cannot prove. "Freshly made", "locally sourced", "hand-cut". Strike them unless the chef will sign for them.
• Heat levels that do not match the kitchen. A model's idea of mild is not your sauce's idea of mild. Use one scale on every dish (mild, medium, hot, very hot) and ask whoever cooks it to confirm each rating.
• Allergen or dietary wording written by the model. Take it from your allergen records, and change the text the same day a recipe changes.
• Ten descriptions that open the same way. "Crispy…", "Juicy…", "Succulent…". Ask for varied openings, or rewrite the first word of any run of three.
• Words your customers do not use. Models drift towards whichever vocabulary they saw most, such as fries for chips or eggplant for aubergine. Put your customers' words in the style sheet.
• Stale text after a recipe change. The description still promises the old sauce. Keep an approval date beside every line.

A monthly check takes 15 minutes. Pick five items at random, compare each description and photo with what is really going into the box, and fix the one that has drifted. Treat any "not as pictured" complaint as a reason to run the check early on that item.

One master sheet so every app tells the same story

Once you list on more than one app, the same dish drifts into three versions of itself. A single master sheet, edited in one place, prevents that and gives the AI one clean source to rewrite from. A filled row looks like this.

The same sheet feeds your counter menu, your website and each app. AI for a cloud kitchen: keep every brand's menu consistent on apps goes further on keeping several menus in step, and the habit pays off the day a supplier change forces a recipe edit and you have to touch four listings instead of hunting for them.

Questions takeaway owners ask before editing a live app menu

Can I put AI-generated food images on a delivery app?

You can make them, but the apps expect a photo to show what actually arrives. DoorDash rejects photos that do not match the listed item, and its photo guidance also names added text and artificial-looking backgrounds among the reasons. An invented dish also invites 'not as pictured' complaints. Use AI to improve a real photo, and read each app's current photo rules before you upload.

How long should a delivery app description be?

About 12 to 25 words, with the contents in the first line, so it reads at a glance on a phone. Check your menu editor for a character limit and give that number to the AI. If a dish needs more explanation, such as a combo with several choices, put the detail into clearly named option groups rather than a longer description.

Should I rewrite the whole menu at once?

No. Change ten items, leave ten similar ones untouched and compare two weeks of sales per 100 orders. Small batches show you what worked, catch wrong claims before they spread and keep the chef's checking time manageable. Start with best sellers that have weak photos or descriptions that only repeat the name.

Does a better menu lower the commission I pay?

No. On DoorDash, for example, commission is a percentage of the order subtotal (15%, 25% or 30% depending on the plan in its published pricing, and rates differ by market), so better copy does not reduce it. The gain comes from more orders, fewer refunds or bigger baskets. Because commission rises with every extra order, check the margin on a dish before you push it.

Further reads

• Restaurant Menu Price Increase: Test Each Change by Dish With AI — Use the same test-and-control method when you change a price.
• AI Ordering vs Delivery Apps: What a Takeaway Actually Pays — Compare what the apps take with what your own ordering channel costs.
• Takeaway EPOS Systems Compared: Which AI Features Matter — See which till features can feed menus to the apps for you.
• AI for Takeaways: What Is Worth Automating and What Is Not — Decide which other takeaway jobs are worth handing to AI next.
• Restaurant Google Business Profile: AI-Drafted Menu, Hours, Photos — Reuse the same photos and wording on your Google profile.
• AI-Generated Marketing Photos: A Cafe Test of What Looks Honest — A cafe test of which AI-edited photos still look honest.
• AI Phone Ordering for Takeaways: A Test Checklist for Menus and Postcodes — Check how a phone-ordering assistant copes with a numbered menu.
• Can a Takeaway Take Orders on WhatsApp With AI? — Move repeat customers to a channel that pays no app commission.
• Kiosks AI for Cafés and Takeaways: Self-Service Ordering Costs — Hardware, software fees, menu set-up and a break-even sum for a self-service kiosk, plus the cases where a person at the counter still wins.
• AI Restaurant Menu Maker: Printable and QR Menus in Canva — Build a printable and QR menu in Canva with AI help, then run a price and allergen check before the print run.
• What Is a Ghost Kitchen? How an Existing Restaurant Can Test One — What a ghost kitchen is, and how a restaurant can test one from its current kitchen on one delivery app: menu, per-order economics and a four-week log.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: DoorDash merchant help and learning-centre pages on photo requirements, rejected photos, the photo editor and AI Photo Enhance, read through search results in October 2026 (the pages disagree on file size and review time, so neither is quoted as fixed); DoorDash merchant pricing page (15%, 25% and 30% delivery commission on the subtotal; other markets differ); Canva pricing page, checked 27 October 2026. Uber Eats photo figures are not quoted because third-party guides disagreed.