How the AI thinks about your menu

When a customer asks the AI “what’s the most interesting vegetarian thing tonight?”, the AI doesn’t read the public menu from top to bottom like a diner does. It assembles an answer from three different layers, in a specific order, and the quality of the answer depends on what you’ve put into each one.

This page is for owners and chefs who want to know what makes the AI sound clever (or generic). After this, you’ll know exactly which fields to fill in to make your AI assistant sound like a senior server who has worked your floor for ten years.

What it does

Think of the AI as having three notebooks open on its lap when a question comes in.

The notebook	What’s in it	Who fills it in	Updated when
The live menu	What’s available tonight, at what price, with which modifiers. Compressed into a single block of text.	You — by editing the menu and POS items.	Every menu or price change.
The catalogue	Every dish’s structured facts: what ingredients are in it, which region it’s from, what allergens it carries, what wine grape it uses, what cooking technique.	The system — automatically, every time you save a menu item.	Within seconds of every menu edit.
The narrative	The chef’s voice. Why this dish is on the menu. How to describe it. Pairing logic. Sourcing stories. House style. Comes from the dish description, the dish’s internal notes, and the venue’s AI context settings page.	You.	Whenever you edit it (which can be never if you’re happy with what’s there).

The first notebook tells the AI what you can sell tonight. The second tells it what those things actually are. The third tells it how you would describe them. Combine all three, and the AI sounds like you.

How to use it

The vegetarian-question scenario

A customer asks: “What’s the most interesting vegetarian thing on the menu tonight?”

The AI’s reasoning, step by step:

From the live menu, the AI pulls the list of every item currently orderable. This morning’s 86’d dishes are already filtered out. The pricing is tonight’s price, not last week’s.
From the catalogue, the AI filters that list to items tagged vegetarian. The tagging happened automatically when you saved each menu item — the system read the description, the ingredients, and the recipe, and applied the dietary classification.
From the narrative, the AI reads your venue’s AI context to learn that “the parmigiana di melanzane is the chef’s mother’s recipe — we only run it when the eggplants from the Friday market are good”. It also reads the dish’s description on the menu item to learn that the “basil is picked the morning of service from the chef’s father’s terrace”.
The AI composes an answer that combines all three: “Tonight I’d open with the parmigiana di melanzane — the chef’s mother’s recipe, the eggplants were picked Friday at the market, the basil comes from the family terrace. We can pair it with a glass of Fiano if you’d like.”

If any of the three notebooks is empty for this dish, the answer gets thinner at exactly that point. A dish with no narrative gets a generic recommendation. A dish with no allergens declared gets recommended to someone it should not be. A dish that the AI didn’t know was 86’d gets recommended even though the kitchen pulled it at noon.

The rule that decides what the AI sounds like

The narrative is your voice; the catalogue is the AI’s memory; the live menu is its truth-check.

If you skip the narrative, the AI sounds like a textbook. If the catalogue is wrong (e.g. you forgot to mark a dish vegetarian), the AI confidently misses dishes that fit. If the live menu is stale (e.g. you didn’t 86 something the kitchen pulled), the AI recommends what’s not coming out tonight.

Each layer protects the AI from a different failure mode.

You don’t have to maintain all three. The system maintains the catalogue automatically. You maintain the live menu the way you already do (the dish editor, the POS price, the availability toggle). You maintain the narrative when you have something to say — not on every dish, only where the description deserves more than the default.

Worked examples

🍕 A dish with all three layers strong — Pizza Margherita. Live menu: priced €13, available. Catalogue: ingredient mozzarella di bufala, region South, technique wood-fired, dietary vegetarian. Narrative: “the mozzarella is delivered every morning from Battipaglia; the basil is from the terrace; we keep an Aglianico open by the glass to pair.” AI output: a recommendation that sounds like the restaurant.
🍅 A dish with only the live menu and catalogue — A spaghetti with tomato sauce, just added to the menu. Live menu: priced and available. Catalogue: auto-tagged Italian, Center, contains gluten, vegetarian. No narrative yet. AI output: “the spaghetti with tomato sauce — Italian, gluten-free pasta available on request.” Functional, but generic. The narrative is where personality enters.
⚠️ A dish with only the narrative — You wrote a beautiful story about the carbonara, but you forgot to mark it as available on the POS, and the system can’t recommend it. The AI never surfaces it to a guest who would have loved it. The live menu is the truth-check — if a dish isn’t there, no amount of narrative makes it reachable.
🍷 A pairing question — The customer asks “what wine with the duck?”. The AI pulls (1) every wine currently available tonight, (2) the grape, region, and tasting profile of each from the catalogue, (3) the venue’s pairing voice (“we pour Nebbiolo with our duck, not Pinot Nero, because the kitchen finishes it with a sour-cherry reduction that wants more acid”). The answer is grounded in tonight’s by-the-glass list and your house style — not in a textbook.

What to write more of, where

If you have an hour a week, here’s where it goes furthest:

Your venue’s AI context (Settings → AI context). 500–1500 words of philosophy, voice, sourcing principles, signature-dish stories. Edited once, helps every dish, every conversation, every customer. Highest leverage.
The dish description on signature items (Edit menu item → Long description). The 5–10 dishes guests notice. Apply the four-part dish pattern (hero → sensory → sourcing → pairing).
Internal notes on the 10–20% of dishes with hidden context (cross-contamination flags, supplier substitutions, upsell rationale). The AI reads these; the public menu never does.

The other 80% of dishes don’t need narrative. The catalogue does enough for them.

Where your information lives — five places, one restaurant — the tables underneath the three notebooks
How the AI uses your food & beverage knowledge — the same idea from the F&B angle
Writing good dish descriptions — the four-part pattern for the dish description
Recommending wine with food — pairing logic the AI applies