AI Menu Descriptions: Writing Restaurant Copy With AI
Short answer AI writes good menu descriptions only from facts you give it. List each dish's ingredients, cooking method, sides and allergens, paste a few lines you already like, and ask for short versions per channel. Then check every word: AI invents sourcing and misses allergens, and a menu is advertising, so false claims count.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callThe short version: facts in, three lengths out, you check
AI can't taste your food. It only knows what you tell it, and fills every gap with the most likely words, which are the same words every other menu uses. Google's Search Central puts the risk plainly:
"Keep in mind that generative models don't retrieve facts, but predict a likely sequence of words based on their training data." (Google Search Central: guidance on using generative AI content)
So the work splits three ways:
| Step | Who does it | Time for a 40-item menu |
|---|---|---|
| 1. Write a fact line for each dish | You or your chef | About an hour |
| 2. Turn the facts into descriptions in your voice | AI | Minutes |
| 3. Check every claim, allergen and price | You, plus whoever cooks the dish | About an hour |
Skip step 1 and you get "a delightful medley of fresh flavors." Skip step 3 and you may print a claim you can't back up.
What to give AI for each dish
The fact line is where the description gets its detail. A spreadsheet with one row per dish works best, because you can paste the whole menu in at once and fix it in one place later.
| Column | Example |
|---|---|
| Dish name as printed | Brisket plate |
| Main ingredients | Beef brisket, house pickles, white bread |
| How it's cooked | Smoked 14 hours over post oak |
| Sides or choices | Two sides; sauce on the side |
| Where things come from (only if true and provable) | Brisket from a named ranch you buy from |
| The nine major allergens it contains | Wheat (bread) |
| Diet tags you're sure of | None |
| Price | $24 |
| One thing staff always say about it | "Ask for the burnt ends if they're left" |
The last column matters most. It's the line only your restaurant can say, and it's what keeps the copy from sounding like every other smokehouse.
A prompt that writes the whole menu in your voice
Paste three to five descriptions you already like (from your current menu, a server's pitch, or a good review), then the spreadsheet, then this:
- "These are descriptions written the way we talk. Match their length, tone and word choice."
- "Write one description per row, using only the facts in that row. If a fact is missing, leave it out. Never add ingredients, sourcing, cooking methods or health claims that aren't in the row."
- "Give three lengths: a menu line under 15 words, a delivery app version under 40 words, and a website version under 60 words."
- "Don't use: delightful, mouthwatering, elevate, perfectly, symphony, medley, culinary journey."
- "Return a table with the dish name first, so I can paste it back into my sheet."
Rule 2 is the one that protects you. For how to save your voice so every chat uses it, see ChatGPT prompts that keep your brand voice.
Any general chatbot does this. ChatGPT's Free plan includes unlimited text chats, with "Limited messages with uploads" (ChatGPT: Plans and pricing), so paste the table as text rather than uploading a file. Dedicated "menu description generators" run the same kind of model; what you pay for is a form and an export, not better facts.
Which words actually make people order?
Specific, appealing words about taste and cooking beat plain names and beat health words. The best evidence is a 2017 Stanford study in JAMA Internal Medicine: over 46 lunch days, a cafeteria randomly labeled the same vegetable dish four ways. The indulgent labels (taste and cooking words, like "sweet sizzlin' green beans and crispy shallots") led to:
"25% more people selecting the vegetable than in the basic condition" (Turnwald et al., JAMA Internal Medicine, 2017)
The same labels beat "healthy restrictive" ones such as "light 'n' low-carb" by 41%, with no change to the food.
You will also see "descriptive labels raise sales 27%." That figure comes from a 2001 study by Brian Wansink, a Cornell researcher who resigned in 2018 after a university finding of academic misconduct; journals retracted several of his papers (Retraction Watch, September 19, 2018). Treat that number with care.
What this means for your prompt:
- Ask for taste, texture and cooking words that are true: "smoked," "charred," "hand-cut."
- Lead with what it tastes like, not what it lacks. "Low-fat" sells worse and triggers rules (below).
- Keep it short. Diners scan; the dish name and the first few words do the work.
Where AI gets menus wrong, and the rules that apply
A menu is advertising. The FTC says so directly:
"Other things like menus, restaurant décor, or even t-shirts and hats worn by servers or customers can be ads too." (FTC Business Blog, September 18, 2024)
That post is about a restaurant whose "Eat local!" menu implied its imported, farmed seafood was caught nearby. Its warning: "False implied claims are illegal." AI loves words like "local," "fresh-caught," "wild" and "farm-to-table" because they appear on so many menus. Delete any you can't prove.
| What AI adds | Rule (US) | Rule (Canada) |
|---|---|---|
| "Local," "wild," "house-made," a farm name | FTC Act: claims, including implied ones, must be true and backed up (FTC blog above) | Competition Act s. 74.01: no representation "false or misleading in a material respect" |
| "Low-fat," "heart-healthy," "high-protein" | 21 CFR 101.10: a restaurant making a nutrient or health claim must provide nutrition information on request, and the dish must meet the claim's definition | Not covered here; check before printing |
| Allergens left out or guessed | FDA's model Food Code (2022) calls for written notice of the nine major allergens in unpackaged food: milk, egg, fish, shellfish, tree nuts, wheat, peanuts, soybeans and sesame. It's a model; your state or county decides whether it's law (FDA) | No source found for a general menu rule; follow your provincial health unit |
| Calories | Chains with 20 or more locations must show calories on menus (FDA: Menu Labeling Requirements) | Ontario: chains with 20 or more locations in Ontario must post calories and a daily-needs statement (Healthy Menu Choices Act, 2015) |
Google's advice for any AI text applies here too:
"It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." (Google Search Central)
Have the person who cooks each dish read its line. They catch "aioli" when it's mayo and "grilled" when it's pan-fried.
Delivery apps write descriptions for you, with limits
DoorDash's merchant tool drafts descriptions from what you've entered, and you can keep, regenerate or edit them. Per DoorDash Merchant Help:
"AI-generated descriptions include a disclaimer until you manually edit the text." (DoorDash Merchant Help: Add menu items)
Anything you type yourself replaces the AI version. The same page bars descriptions that mention other delivery apps, steer customers to your own site with a lower price, or refer to delivery fees or commissions. So don't paste your website copy in unchanged if it says "order direct and save."
Your website and Google Business Profile should carry the same text, as real text rather than a PDF image, so search and AI assistants can read it; see AI search for restaurants.
Example
Illustrative cases, round numbers.
A taqueria with 30 items in San Antonio
The owner filled the sheet in 45 minutes, pasted four lines from her chalkboard specials, and got all three lengths in one chat. AI wrote "locally sourced corn tortillas" for a dish where the tortillas come from a distributor; the "never add sourcing" rule was missing from her first prompt. She added it, reran, and the cook caught one wrong salsa. Total time: about two hours, $0.
A Toronto café with 4 locations
The owner asked AI to add "guilt-free" and "high-protein" to the bowls. He dropped both: he had no nutrition numbers to back them, and the descriptions read better with "roasted squash, tahini, crunchy chickpeas." With 4 locations, Ontario's calorie posting rule doesn't apply to him; at 20 it would.
A pizzeria updating DoorDash
The DoorDash AI drafts were accurate but bland and carried the disclaimer. The owner pasted his own 40-word versions, which removed the disclaimer, and deleted "order on our site for 10% off" from two of them.
FAQ
Is it OK to use AI for restaurant menu descriptions?
Yes. Google doesn't ban AI text; it warns against mass pages that add nothing. You answer for every claim, so check each line. See does Google penalize AI content?
How long should a menu description be?
Short enough to scan: under 15 words for a printed menu. Delivery apps and websites take more, so ask AI for each length separately.
Can AI write allergen information for my menu?
It can format what you give it, but never let it decide what a dish contains. Allergens come from recipes and supplier labels.
What's the best free AI menu description generator?
Any general chatbot's free plan works if you feed it a fact sheet. Paid generators add templates, not knowledge of your food.
Different for you?
- You want copy beyond the menu (About page, ads, emails): see AI copywriting.
- Your menu isn't showing in Google or ChatGPT answers: see AI search for restaurants and Google Business Profile optimization.
- You run 20 or more locations: calorie posting applies in the US and in Ontario; check the full rule before reprinting.
- You'd rather have the menu page written and kept current for search: SEO & GEO Agent, or for the page itself, Website Agent.