The Restaurant Facts Assistants Read About You

What this page claims and what it does not

This is not an introduction. It is the border of what we know, and it comes first on purpose, because on this subject the honest border is narrower than the sales pitch.

We have not tested what any assistant answers about any venue. We do not measure assistant output, we do not track it, and nothing on this page is derived from it. Every formula below measures something else entirely: what the restaurant itself published, and whether those publications agree with one another. That is a different object of measurement, and the difference is not a nuance — it is the whole design of the page.

There is also no official statistic to lean on here. The Eurostat catalogue lists 6 653 datasets, 215 of them in the isoc_* family that covers digital use by enterprises and households; a sweep of the whole table of contents for review, rating, opinion, feedback, assistant, chatbot, visibility and listing returns 35 lines, and every one of them is about asylum appeals, life satisfaction ratings or employee feedback (Eurostat, catalogue table of contents, opened 26.08.2026). No dataset describes what assistants answer about businesses, and none describes how many people read business listings before choosing a venue. So this page names no benchmark for either. It gives arithmetic instead.

Read the right-hand column as the point of the page, not as a disclaimer stapled to the bottom of it. Everything we sell on this subject lives in the left-hand column.

The source of record: the one place a fact about the venue is true

Source of record — the one place the restaurant declares a fact to be true. Without it, consistency cannot be measured, because there is nothing to be consistent with.

Published fact — a statement about the venue that exists outside the restaurant: hours, address, phone, menu, prices, booking and deposit rules, accessibility, allergens.

Most venues have the second without the first. Hours live in four places, three of them edited by someone who left, and when they disagree there is no way to say which one is wrong — only that they differ. Naming a source of record turns an unanswerable argument into a checkable one.

Which place you pick matters less than the fact that you picked

The source of record can be your own site, your reservation system, or the Google profile — the choice is yours and it is a management decision, not a technical one. What is not optional is that the choice exists, is written down, and is known to everyone who edits anything. A restaurant that says "the site is the truth for hours and the reservation system is the truth for the deposit" can measure everything below. A restaurant that has not decided cannot start.

The order of filling those places in — who edits what, and in what sequence — is held by a separate article: restaurant data in Google, on the site and on portals. This page does not repeat it. This page measures the result.

The eight facts most often checked about a venue

The list below is a working table, not a benchmark. Fill the second and third columns yourself; the fourth is what tells you where contradictions will appear first.

Eight rows is the working size: short enough that somebody actually fills it in, long enough that the two facts which cause most of the trouble — exception hours and prices — are both on it. Add rows for what your venue really has (a second entrance, a seasonal terrace, a dog policy) and the arithmetic below does not change.

Consistency: the formula, and why one contradicting source costs more than ten agreeing ones

Consistency rate — the share of checked sources where a fact matches the source of record.

Contradiction — a source that states the fact differently. One contradiction deserves more attention than ten agreements, because a reader who meets it stops trusting all of them.

Consistency rate % = Sources where the fact matches the source of record ÷ Sources checked × 100

• Sources where the fact matches the source of record — count of places whose value is identical to the declared truth, pieces;
• Sources checked — the list of places you wrote down before you started, pieces. Without a written list the denominator changes between measurements and the number stops meaning anything;
• the result is a percentage of sources, not of guests, not of searches and not of answers.

Worked example. You wrote down eight places where your hours appear: your own site, the Google profile, two delivery aggregators, a city guide, a booking portal, the Facebook page and a printed-menu PDF that is also online.

Checked in reverse: 0.625 × 8 = 5, the five that match. Three do not, and those three are the whole job.

The asymmetry in the second definition is the reason this indicator is worth computing at all. A guest does not average your sources. A guest sees one of them, arrives at a closed door, and the other seven agreeing perfectly do not repair the evening. That is why the next section counts contradictions as objects rather than as a rate.

Machine-readable form: what separates a structured fact from the same fact in a picture

A fact printed inside a photograph of a chalkboard is published for a human and invisible to anything that reads text. A fact written into markup is a labelled value with a defined meaning. The vocabulary that defines those meanings is public and can be quoted precisely, which is exactly why this page relies on it instead of on claims about what machines do with it.

The Restaurant type carries, among others, acceptsReservations — "Indicates whether a FoodEstablishment accepts reservations. Values can be Boolean, an URL at which reservations can be made or (for backwards compatibility) the strings Yes or No", and hasMenu — "Either the actual menu as a structured representation, as text, or a URL of the menu" (schema.org, type Restaurant, opened 26.08.2026). Alongside them sit openingHoursSpecification, priceRange, servesCuisine, telephone and address. That is the standard talking about fields, not us talking about outcomes.

What the sector actually publishes is measurable, and here official statistics do exist.

The cut matters more than the figures: section I is accommodation and food service together, hotels included, and only enterprises with at least ten employees. There is no separate restaurant line in this dataset, so none of these numbers is a statement about restaurants as such.

What you can verify about your own markup

Three things, all of them checkable in an afternoon and none of them a claim about assistants: that the field exists in the page source; that a structured-data validator parses it without errors; that the value inside the field equals the value in your source of record. The third is the one that usually fails, because markup is written once and hours change afterwards. Publishing the facts on a site you control is the part you own end to end — see website and published facts.

Opening hours that change: the most frequent source of contradictions

Hours are the fact that breaks first, because they are the only one on the list that has planned exceptions. The standard has a field for exactly that. OpeningHoursSpecification holds opens, closes, dayOfWeek, and the pair validFrom and validThrough — the second one being "The date after when the item is not valid. For example the end of an offer, salary period, or a period of opening hours" (schema.org, type OpeningHoursSpecification, opened 26.08.2026). The same page states the rule that catches most late-night venues out: "The place is open if the opens property is specified, and closed otherwise. If the value for the closes property is less than the value for the opens property then the hour range is assumed to span over the next day."

A venue closing at 01:00 and writing closes: 01:00 with opens: 17:00 is, by that rule, describing a range that spans midnight — which is correct. The same venue writing 01:00 as the closing time of the next calendar day has published two overlapping ranges, and every copy of that mistake is a contradiction.

Worked example, hours only.

Keeping one change propagated across the places that hold it is what automated updates are for, and the profile that most often holds the second copy is covered by the Google business profile.

Menu and prices: why their copies go stale fastest

The menu is the fact with the shortest half-life and the largest number of copies. It is printed, photographed, uploaded as a PDF, retyped by two aggregators and quoted in a review from three seasons ago. The standard permits all three forms — "the actual menu as a structured representation, as text, or a URL of the menu" — and only the first two are readable as values rather than as pictures.

Prices are worse, because a price copied into a third-party listing is edited by that third party, on their schedule, and nobody tells you when it drifts. This is why the price row in the table above is the one where a written source of record earns the most: without it, the argument about which price is real has no referee.

There is no honest way to say how much a stale price costs you, and this page does not invent one. What it can say is arithmetic: every extra place holding a copy adds one to the denominator of the consistency rate and one potential contradiction to the counter, and copies created by typing the same value twice are the cheapest ones to remove — that is what integrations address.

Third-party listings where your data lives without you

Some places holding your facts were never filled in by you at all: aggregators that scraped a menu, city guides built from an old database, map entries created by a guest. You cannot delete them by decision. You can list them, and a listed source is countable.

Reviews are published facts you do not write

Reviews belong to the same set of published facts, and unlike menus they are written by other people while you answer under your own name.

Again: section I is accommodation and food service together, hotels included, enterprises with ten or more employees only. Who writes the answers and on what cadence is the subject of responding to Google reviews, and the operational side of collecting them is reviews and replies.

A chain multiplies all of this by the number of locations, and multiplies the confusion too, because the same brand name with three addresses invites listings to blend. Running three sets of published facts from one screen is described in managing a restaurant chain from one screen, and the rating arithmetic that sits on top of the review pile has its own page, restaurant rating math.

Coverage: how many facts from the list are published at all

Coverage — the share of the facts on the checklist that are published in a machine-readable form at all.

Coverage % = Facts published in machine-readable form ÷ Facts on the checklist × 100

• Facts published in machine-readable form — count of facts that exist as a value in markup or in a listing field rather than only inside an image or a PDF, pieces;
• Facts on the checklist — the length of your own list, pieces. Ten rows or eight, the denominator is whatever you wrote;
• the result is a percentage of facts, and it says nothing about how anyone uses them.

Coverage moves slowly, and that is the point

Worked example. Your checklist has ten rows; six of them exist as values — address, phone, regular hours, reservation rules, price range and cuisine.

Checked in reverse: 0.6 × 10 = 6. There is no norm attached to that number, and this page refuses to invent one: no official source publishes a coverage benchmark for restaurants, so the only meaningful comparison is your own figure a quarter later against your own figure today.

Contradictions are counted in pieces, not in percent

Contradictions = Σ Sources stating the fact differently

• counted separately for each fact on the checklist, pieces;
• it is a counter, not a share, and it must not be folded into a percentage.

Worked example. Hours disagree in two places, the phone number in one, prices in three: 2 + 1 + 3 = 6 contradictions across the whole checklist. The degenerate case is worth printing rather than hiding: a rate would tell you that six contradictions out of forty checks is "fifteen percent", which sounds tolerable and is meaningless — the guest who reaches the one wrong entry is not consoled by the thirty-four correct ones. Each individual listing matters, so each is counted as an object, and the work list is the six, not the percentage.

Where AI works here and where it is plain arithmetic

The three formulas above are arithmetic. Division and addition, computable in a spreadsheet, no model involved, no judgement. If a product tells you it "measured your visibility", ask which of the three it computed, and on which written list of sources.

Where a model can genuinely help is upstream of the arithmetic: reading unstructured text from a listing and proposing which value it thinks corresponds to which field, so a human confirms or rejects it rather than retyping. That is extraction, and it is checkable — the human sees both the source text and the proposal. The management layer that stores the confirmed values and keeps them in one place is described in an AI restaurant management system, and the wider operational picture in restaurant automation.

Section I again: accommodation and food service together, hotels included, ten or more employees.

The honesty boundary: what this work does not give, and why no timeline appears here

Three sentences that do not get softened.

First, nobody can promise that an assistant will describe a venue in a particular way, and this page does not attempt it in either direction — not as a promise, and not as a denial dressed up as expertise. The mechanism is not published by anyone whose statement we could open, quote and date, so we say nothing about it.

Second, no timeline appears on this page. Not for markup, not for reconciliation, not for any effect. A date is the easiest thing in the world to write and the hardest to be held to, and a date placed in an answer that machines quote becomes a lie that repeats itself.

Third, the work has a shape and the shape is small: a written list of sources, a written source of record, three numbers recomputed on the same list, and a shrinking list of contradictions. Everything in that sentence is yours, verifiable by you, and unchanged by whatever any model does downstream. The rest of the operating picture — where these facts live inside the business rather than outside it — is the restaurant economics hub, and the queue that stops the same fact being typed twice is one queue instead of five.

Frequently asked questions

Where does an assistant get information about a restaurant?

From text that already exists somewhere public: the venue's own site, its listings, aggregators, maps, reviews, articles. None of that is under a single owner, and we have not measured which of it any assistant uses. What is measurable from your side is the other end of the same pipe — what you published, in what form, and whether the copies agree.

What is a source of record, and why does a restaurant need one?

It is the one place where the restaurant declares a fact to be true. It is needed for a mechanical reason before any marketing reason: consistency is a comparison, and a comparison needs something to compare against. Without a declared source of record you can observe that four listings differ, but you cannot say which of them is wrong.

How do I measure whether my published facts agree with each other?

Write down the list of places to check, then for each fact divide the number of places that match your source of record by the number of places checked and multiply by a hundred. Five matching out of eight checked gives 62.5 %. The written list is the load-bearing part: change the list between measurements and the two numbers are not comparable.

Does structured data change what an assistant says?

We do not know, we have not tested it, and no source we could open, quote and date establishes it — so this page makes no claim either way. What the standard does define is verifiable and quotable: which fields exist, what each one means, and what values it accepts. Publishing a fact as a labelled value rather than as a picture is a decision about your own data, and it stands on its own merits.

Which facts change most often and create contradictions?

Exception hours and prices, by a wide margin, because both change on a schedule set outside the marketing calendar and both are copied into places you do not edit. Menu composition follows. Address, phone and accessibility change rarely, which is why a contradiction in one of them usually means a stale listing nobody has touched in years rather than a recent mistake.

Can a restaurant control how an assistant describes it?

No, and anyone saying otherwise is selling something they cannot deliver. What a restaurant controls is its own publication: whether each fact exists in machine-readable form, whether every listing it can reach carries the same value, and how quickly a change reaches those listings. Those three are measurable and improvable. The description itself is not in anyone's hands.

Where is AI at work on this page, and where is it plain arithmetic?

All three formulas are arithmetic — two divisions and one sum, reproducible in any spreadsheet. A model can be useful before the arithmetic starts, by reading a listing's free text and proposing which value belongs in which field for a human to confirm. Nothing on this page uses a model to produce a number, and nothing here measures a model's output.

Take the checklist, write down the places where each fact appears, and compute the three numbers once. The contradictions will be in the hours; they always are, and they are found with a table rather than with technology. When you want the same reconciliation run as part of a management layer rather than by hand, start from the restaurant economics hub.