AI content strategy by dining occasion: the mistakes that empty your calendar and the method that fills tables

Step 1 · Pull your POS into an eight-slot grid before writing a single line

Export twelve months of tickets and group them into eight time slots: that file is the deliverable of the first step, and without it there is no strategy anyone can defend. Every row carries five measured columns, never estimated ones: average covers, average ticket, minutes of dwell time, heaviest day, and dominant reason for the visit. Verification is simple, and I suggest you do it with your accountant sitting there: the covers across all eight slots must reconcile with annual sales on the P&L within a margin under 2%. If it doesn't reconcile, the problem lives in the POS and not in marketing, and fixing that first saves you six months of content aimed at the wrong place. Menu prices at large U.S. chains climbed 42% between 2020 and 2025, nearly double the 22% general inflation, per One Haus; under that ticket pressure, publishing without knowing which slot can absorb an increase is gambling.

Step 2 · Attach one pain and one cash figure to every row of the grid

A slot with no named pain is not yet a brief, and that is the deliverable of the second step: eight sentences, one per row, stating what hurts the person seated there. The two-thirty lull does not carry the same problem as the office lunch rush; a table that holds ninety minutes and leaves 96 USD across four people needs reasons to stay, while the 45-minute, 32 USD table needs reasons to come back on Thursday. Write the pain in the customer's words, not in yours. And anchor the number: what share of weekly sales that slot contributes today, and what the ninety-day target would be. With digital orders up 237% since 2020 in full-service restaurants, per Restroworks, several of those eight rows no longer happen inside the dining room at all. A useful prompt is a situation with rules, not a topic.

Step 3 · Turn each row into a prompt with hard constraints, not a polite request

The model needs the whole slot handed to it —hour, ticket, minutes, pain, target— plus an explicit ban on whatever ruins the output: no promises the kitchen can't hold, no prices unverified against the recipe costing, no offers that don't exist on the menu. Diego F. Parra insists that Masterestaurant works backwards from how the AI courses teach it: the grid first, the tool second, because an unconstrained model returns the industry average and the industry average will not fill your Tuesday. The deliverable here is eight prompts stored in a living document, versioned with dates. When a dish's food cost climbs past the 32% ceiling we set, you edit one prompt instead of forty pieces. Generate five pieces per slot, forty in all, and do it in batches separated by row: mixing slots in one session contaminates the output and hands you forty variants of the same text.

Step 4 · Produce in batches by slot and verify uniqueness before scheduling

Verification has two parts and neither is optional. First, read two pieces from different slots out loud; if you could swap them without anyone noticing, the grid never made it into the prompt. Second, measure: no pair of pieces should share more than 45% of its vocabulary. That threshold is not an editorial whim, it is survival, because Google has penalized template pages that change one variable since 2026, with documented traffic drops reaching 80%. Production today costs about what lunch for two costs. The judgment to decide which forty still costs exactly what it always did. Let me start with the most expensive one: reversing the order and buying the tool before the table exists. Three more follow behind it. Publishing eleven times a week while ten of those posts talk about the same signature dish —we saw this at a 180-cover restaurant in Bogotá that filled every Friday while two-thirty sales ran 38% below the weekly average— leaves a full calendar and an empty Tuesday register.

The four mistakes that wreck execution, ranked by how often they show up

Measuring reach instead of covers turns the report into decoration nobody can argue with. And hand-editing one piece when the output fails, rather than editing the prompt that produced it, sentences you to repeat that correction forty times. The deliverable of this step is one sheet listing those four mistakes with the early warning sign of each, pinned where the marketing team can see it. Every Monday, one single table: the week's posts by slot against actual covers in that same slot. That is where marketing stops being an opinion. If the afternoon lull got five pieces and its covers haven't moved in three weeks, the problem is not frequency but the pain you named badly back in step 2, and you go back there before producing anything else.

Step 5 · Cross publishing against covers and let the slot speak in numbers

Signals arrive on uneven delays: personalized email lifts opens by 26%, per Stripo, and shows up within days; a creator post moves reservations by as much as 30% the following week, per Marketing LTB, visible in seven days; one additional review star is worth 5% to 9% of revenue, per Michael Luca's study at Harvard Business School, and that takes quarters. Measure each hypothesis on its own clock. Suppose you ignore everything above and produce the forty pieces on a Friday afternoon with a good model and good taste. Week one, the team celebrates the volume. By week three, somebody asks for the report and all that exists is reach, because there is no slot to cross it against. In month two, the owner notices Fridays are still packed and Tuesday is still empty, and concludes AI doesn't work for his restaurant. By month four, he cancels the tool and goes back to an agency that will charge eight times more to produce half as much.

What happens if you skip the grid and go straight to the generator?

The mistake was never the model, which did exactly what it was asked. It was asking it to segment without giving it anything to segment against, and the whole business pays that invoice for a year.

You are finished when you can tick these seven, and not before. One: the eight-slot table exists and reconciles with the P&L under 2% deviation. Two: every row carries a written pain and a ninety-day sales target. Three: eight date-versioned prompts live in a working document. Four: the forty pieces sit five per slot, with no convenient exceptions. Five: no pair shares more than 45% of its vocabulary. Six: the Monday board crosses publishing against covers, not against reach. Seven: each slot has a named owner. The definitive test takes one question, and run it cold a month from now: grab a piece at random, cover the headline, and ask your floor manager what hour it is talking to.

Closing checklist · the seven boxes that confirm it landed

If the answer comes back right, the strategy is alive. Sequence. Grid first, prompt second: invert that order and you end up with forty beautiful pieces that speak to nobody at any particular hour. The grid comes out of the POS in one afternoon and fits in a single eight-row table with slot, average covers, check average, dwell minutes and dominant visit motive. Without that document, artificial intelligence for restaurants has nothing to segment against, so it returns the industry average: correct, generic, useless. Where the metric lands. A calendar that dies at reach is a calendar nobody can defend to the owner. Cross Tuesday's post with Tuesday's covers and marketing becomes decision intelligence: each cluster carries a hypothesis (add 12 covers to the soft slot), a deadline (three weeks) and a POS verification. When the hypothesis fails, you change that occasion's promise rather than the photo.

The three differences that decide the outcome

Reuse. Twelve dish posts expire in 48 hours; eight occasion clusters live nine to twelve months, because Thursday after office will still be Thursday after office in March and in October. That compounding return is what most operators leave on the table: one production effort pays out across four or five cycles, and AI agents only need refreshed figures and seasonality to send it live again.