Is your restaurant visible in AI search? Run the four-number check.

Stand at your pass on a slow afternoon, pull out your phone, and ask it out loud for a good version of what you serve, near you. Don’t scroll — just listen. You get three names. Maybe five. Not ten blue links, not a map with twenty pins — three names, read off the top of the answer. Now the only question that matters for the next hour: was one of them yours? If it wasn’t, the diner never learned you exist, and your hard-won position on the old Google map never got read aloud.

That ten-second test is the whole game in miniature, and most operators have never run it on their own listing. To appear in AI search, a restaurant has to clear four numbers — its star rating against the assistant’s floor, review recency, profile completeness, and whether its own site holds an answer the AI can quote — and you can read all four this week without a consultant. A 2026 Uberall study of multi-location restaurants found that 83% never appear in AI search results, even though 86% of them keep a Google Business Profile. Having the listing is not the same as being named. The assistant decides who makes the shortlist before ranking ever enters the picture — and this is the field guide to reading the four numbers that decision turns on.

You do not need a consultant to find out which side of that 83% you are on. Four numbers decide it, and you can read all four this week.

Why presence stopped being enough

When a diner asks ChatGPT, Gemini, or Google’s AI Overview for a recommendation, the assistant does not return a ranked list of everyone nearby. It returns a written answer that names three to five places. Uberall measured the shortlist at three to five brands per query. Everyone else is not on page two — they are simply not in the sentence.

So the goal moved. Ranking first in the map pack still matters, but it no longer guarantees the assistant repeats your name. The assistant filters on a handful of signals first, then writes. Four of those signals are numbers you control — so the rest of this guide is a runbook for reading them, in order, on your own listing.

You are not competing for a rank anymore. You are clearing four bars before the assistant will say your name out loud.

Number one — your rating against the floor

Each assistant appears to apply a minimum star rating before it will name a place at all. Uberall’s 2026 figures put the floors at roughly 4.3 for ChatGPT, 4.1 for Perplexity, and 3.9 for Gemini. Below the floor, completeness and rank do not save you — you are filtered out before the assistant starts choosing.

The move is not “get more stars.” It is to know your current number and the gap to the nearest floor. A restaurant sitting at 4.0 is invisible to ChatGPT’s shortlist no matter how good the food is; the same restaurant clears Gemini’s 3.9 today. Pull your number, write down the gap, and you know exactly which assistant you are losing. The GBP grader reads it for you if you would rather not do the arithmetic by hand.

Number two — review recency, not just the count

A 4.6 built on twelve reviews, the newest from eighteen months ago, reads as a closed restaurant to a model trained to weight freshness. A 4.4 with a handful of reviews this month reads as open, busy, and current. Count earns the attention; recency keeps it.

The lever here is review responsiveness — fresh reviews arrive faster when you are visibly answering the ones you already have. The full playbook is in how to respond to Google reviews; the short version is that a reply is a dated signal of life, and the assistant can read the date.

Number three — how complete the profile actually is

86% have a profile. Far fewer have a profile the assistant can actually read. A listing with no hours-by-day, no cuisine set, no service attributes, and no menu link is a thin signal — the model has nothing to quote when the diner asks “are they open Sunday” or “do they do gluten-free.” Presence, again, is not the same as legibility.

Fill the structured fields on your Google Business Profile the way you fill out a permit: every field, no blanks, current. Hours by day, cuisine, dietary attributes, outdoor seating, the link to your real menu. The GBP grader flags the empty fields; the SEO grader checks whether your own site backs them up.

Number four — do you own an answer the AI can lift?

The assistant needs a sentence to quote. If the answer to “is it good for a group of eight,” “is there a vegan entrée,” or “can you book a patio table” lives only inside a photographed menu or a PDF, the model cannot lift it — it quotes a competitor who wrote the answer in plain text. The restaurants that get named are the ones that put the answer where a machine can read it.

That means text on your own pages plus restaurant schema markup so the structured facts are unambiguous. Run schema check to confirm the markup parses. This is the one number that lives on your site rather than your Google listing, and it is the one most operators have never looked at.

The fixes that backfire

Two reflexes make the numbers worse. Buying reviews spikes the count and then tanks recency the moment the batch ages, and Google’s filters increasingly strip them — so the average drops and the dates go stale at once. Stuffing keywords into your business name is a suspension risk that can pull the whole listing, taking all four numbers to zero in a day. The signals that move AI search are the slow, real ones: good food showing up as fresh reviews, a complete profile, and answers written in plain text.

The ten-minute version

You do not have to fix all four today. You have to read all four today, so you know which one is costing you. Here is the field check, in order — open your listing on your phone, set a ten-minute timer, and write the four answers on whatever’s nearest.

1. Read your rating against the floors. Open your Google listing and note the star number to one decimal. Subtract: how far above — or below — 3.9 (Gemini), 4.1 (Perplexity), 4.3 (ChatGPT)? Write the gap to the nearest floor you’re under. That gap is the one assistant you’re losing.
2. Check the date on your newest review. Not the count — the date. This week? This month? Or has it been a season? A stale top-of-list reads as a closed kitchen no matter how high the average. Write the month of the most recent one.
3. Walk the profile for blanks. Hours by day, cuisine, dietary attributes, outdoor seating, the menu link — tap through each. Every empty field is a question the assistant can’t answer about you. Write down how many blanks you find.
4. Search your own site for the answer. Pick a real question a diner asks — “vegan entrée,” “group of eight,” “patio booking” — and find where it lives. If the answer is only inside a photo or a PDF, mark it “not extractable.” That’s the number most operators have never checked.

Four answers on a scrap of paper, and you know which number is keeping you out of the answer. Close the nearest gap first — the one that needs the smallest move to clear a floor or fill a field.

The fourth number is the one that decides number four, so it is worth seeing the difference up close. Here is the same fact — “can you seat a group of eight on the patio” — written two ways. The first is how most restaurants ship it; the second is what the model can actually quote.

Four numbers, ten minutes, no agency. The 83% are not invisible because AI search is unfair to small restaurants. They are invisible because no one has read these four numbers and closed the nearest gap.