83% of Restaurants Are Invisible in AI Answers. Winning the Answer Is the Easy Half: OpenTable Takes the Booking, DoorDash 15–30% of the Order.

Uberall counted 83% of restaurant locations as entirely invisible in AI recommendations. The ones that do get named have a different problem.

Start with the number. On May 7 2026, in a report released through BusinessWire, Uberall found that 83% of restaurant locations are entirely invisible in AI-generated recommendations (Uberall via BusinessWire). Uberall sells location-marketing and AI-visibility software, so a finding that restaurants are invisible in AI sells the product that fixes it. The release does not fully disclose its sample. Read it as the shape of the problem, not a measurement of your block.

It is consistent with the wider local picture: SOCi's 2026 Local Visibility Index (roughly 350,000 locations; SOCi sells local-marketing software) found ChatGPT recommends a given location in 1.2% of answers and Gemini in 11%, against 35.9% for Google's local 3-pack (SOCi). Getting named is rare everywhere.

But the 83% describes only the losing side. It says nothing about the minority of restaurants the assistant does name — the part nobody in the AI-visibility business discusses, because it does not end when you win.

Here is the shape of it. A guest asks an assistant where to eat. The assistant names your restaurant: the recommendation is won, and you paid nothing for it. Then it offers to book the table, and the booking runs through a reservation platform. Or the question was delivery, and the answer ends inside an app carrying a published commission. You won the recommendation; somebody else took the transaction attached to it — the booking, the margin, the record of the guest. The unit of loss inside an AI answer is not the guest. It is the transaction the guest was carrying. And notice where the handoff comes from: the answer routes into the product that holds a machine-readable menu, a machine-readable table and a machine-readable price. If your own site holds none of those in text, the aggregator is not stealing the transaction — it is the only party in the conversation equipped to complete it.

OpenTable is OpenAI's first restaurant tech partner for Operator — the agent that books the table, not merely the one that names it

OpenTable states that it became "the first restaurant tech partner for OpenAI's Operator tool", the use case being Operator helping diners book single and recurring reservations online (OpenTable). Operator is OpenAI's agent, announced January 23 2025, which carries out tasks in a browser on the user's behalf (OpenAI).

Be precise, because the trade press has already blurred it. It is a booking-agent partnership: an agent that can complete a reservation, pointed at a reservation platform. It is not a statement that ChatGPT's restaurant recommendations are powered by OpenTable. No study shows a reservation platform's inventory decides which restaurants an assistant names. Ask any vendor claiming otherwise for the study — there isn't one.

The distinction leaves the problem intact. Once an assistant can act rather than answer, the last step of the decision — where a diner becomes a booking — happens inside a product you do not own. The recommendation is the part you influence, with your own menu, pages and reviews. The transaction is the part handed off, and what the platform passes back to you is whatever your agreement says it passes back. Read yours.

We have no verified figure for what a reservation through a booking platform costs an American restaurant, and we will not invent one. What is published, by the platforms themselves, is the delivery leg.

DoorDash publishes 15%, 25% and 30%. Uber Eats publishes 20%, 25% and 30%. That is the toll on a customer the assistant already sent you.

DoorDash's merchant pricing page lists three delivery commission tiers: Basic at 15%, Plus at 25%, Premier at 30% (DoorDash). Uber Eats' merchant pricing page lists marketplace fees of 20% for Lite, 25% for Plus, 30% for Premium (Uber Eats). These are not agency estimates. They are published rate cards, printed by the companies collecting the money.

Put them in sequence. A guest asks an assistant for late-night tacos. The assistant names your restaurant. The guest says "order it", and the answer routes into a delivery app. You could not have bought that recommendation. But the order arrives through a channel with a published 15% to 30% commission on it, carrying a customer the assistant had already decided to hand you. Between an answer ending on your own ordering page and one ending inside an aggregator sits the difference between two margins: same guest, same night, same food.

What nobody can tell you is how often that happens. There is no current, open, US source putting a share on restaurant orders flowing through DoorDash, Uber Eats and Grubhub. The number does not exist, so a vendor quoting you one is quoting something they cannot show you. What is measurable: across the answers that name your restaurant, how many end with a route into a platform, and how many with a route to you.

A menu locked in a PDF cannot answer the question the guest actually asked

Guests do not ask assistants for restaurants. They ask for a table that solves a problem: gluten-free Italian genuinely safe for a celiac; a room quiet enough to close a deal over lunch. Each is a question about the contents of a menu and the character of a room.

In a great many restaurants, both live inside an image: a PDF menu, a photo of a chalkboard, an Instagram carousel. To a model reading the web, a picture of the words "gluten-free" is not the words "gluten-free" — it is a picture. A restaurant that spent four years building a celiac-safe kitchen has, as far as the assistant is concerned, published nothing about it. It is not losing that answer. It is not in the running for it.

The fix is unglamorous and nearly free: put the menu on your own site as readable HTML text, and write the words a guest uses in the question next to the dishes they apply to. There is no hack underneath it. The peer-reviewed GEO study (Aggarwal et al., ACM SIGKDD 2024, 10,000 queries) found citing sources lifted a page's visibility in generated answers by up to 115% for low-ranked content, statistics by 41% and quotations by 28%, while keyword stuffing produced no gain (arXiv). Checkable writing is what these systems reward, and a menu is the most checkable writing a restaurant owns:

• Dietary and allergen language on the dish itself: gluten-free, celiac-safe, dedicated fryer, nut-free, dairy-free, vegan, halal, kosher. Not a legend of icons — a picture of a leaf is not the word "vegan".
• Occasion language on the page: quiet at lunch, private room, seats 12, patio, high chairs, walk-ins after 10pm, corkage. These are the words inside the question.
• Hours that are true, including the kitchen's last order — which is not the bar's closing time. "Late-night tacos after 11pm" is a question about the kitchen. A listing saying 1am when the pass shuts at 10:30 turns the answer that names you into a complaint.
• Every fact a guest has to phone to find out, because a fact you make someone call for is a fact the engine guesses. SOCi's 2026 index (350,000 locations; sells local-marketing software) put business-data accuracy at about 68% for ChatGPT and Perplexity, against 100% for Gemini (SOCi).

The sources an assistant reaches for when someone asks where to eat, ranked — and what has actually been measured for restaurants (almost nothing)

Nobody has published a per-restaurant citation-share study. The ranking below is built from local-business-wide measurements plus the structure of the restaurant answer itself, and each line says which. Where a line is a hypothesis, log it as one. One line is not a hypothesis: the engine reads text, not your POS and not your reservation book, so a dish is celiac-safe to a model only because some page says so in words. Your menu page is where that sentence can exist; everything under it either repeats the sentence or contradicts it.

The nearest primary evidence is ours, and it is from other verticals. In the three markets we have measured — dental, med spa and plastic surgery in Dallas, July 2026, across ChatGPT, Gemini and Perplexity — 73% to 79% of every source cited behind those recommendations was a business's own website rather than a directory (the Dallas benchmark). Restaurants were not in that sample, so treat it as direction, not as your number. Dallas, July 2026 — one market, one day, and a primary measurement: the raw answers are stored, the method is published, and the run repeats against any city, including yours.

One general finding frames the list. BrightLocal's Local Consumer Review Survey 2026 (n=1,002 US consumers; BrightLocal sells local SEO and reputation tools) found 45% of consumers now ask AI to recommend a local business, up from 6% a year earlier — putting AI ahead of Yelp and TripAdvisor as a channel (BrightLocal). A strange sentence for a restaurant to sit with, because Yelp and TripAdvisor are two of the sources the AI reads to get there.

And one filter sits above everything. SOCi's 2026 index found the average rating of locations ChatGPT recommends is 4.3 stars, Perplexity 4.1, Gemini 3.9 (SOCi; sells local-marketing software) — local business, not restaurants, but a restaurant at 3.8 on Google should treat that as the cheapest hypothesis here.

• 1. Your own menu page, in readable HTML — the only source here where "celiac-safe" and "seats 12" can exist as words rather than be echoed from somewhere else. A listing can carry the fact; it cannot originate it. Our own Dallas run points the same way in adjacent verticals: 73% to 79% of cited sources were businesses' own sites across dental, med spa and plastic surgery (the Dallas benchmark; our primary measurement — one market, one day, no restaurant cut). Yext's October 2025 analysis of 6.8 million citations across 1.6 million AI answers found Gemini takes 52.15% of its citations from brands' own sites, ChatGPT the near-mirror at 48.73% from third-party listings (Yext); Yext sells listings management, so a finding that listings decide everything sells listings management. Measured across brands, never restaurants. Do this: the menu in crawlable text, with the dietary and occasion words on the dish.
• 2. Google Business Profile and Maps — the strongest third-party evidence here, none of it restaurant-specific. It was strongly preferred by Google's own models in BrightLocal's July 2025 test across four engines (BrightLocal; sells local SEO tools), and Gemini's data accuracy hit 100% in SOCi's index because it is grounded in Google Maps (SOCi; sells local-marketing software). Do this: right cuisine category, real hours including the kitchen's last order, every attribute you offer.
• 3. Yelp and TripAdvisor — the biggest measured third-party number in local AI, with no restaurant breakout published. Yelp drew 512,680 citations across 28 million AI answers in Q4 2025 (Foundation Inc. with AirOps, a content agency working with an AI-search vendor, so their own tooling is the subject). Do this: claim both, and make cuisine, hours and attributes match your site word for word.
• 4. OpenTable and Resy — untested as citation sources, and the one place here where the assistant can act rather than answer. OpenTable is OpenAI's first restaurant tech partner for Operator, for booking single and recurring reservations (OpenTable; OpenAI). Ranked for what it does to the transaction, not for a citation share nobody measured. Do this: keep availability accurate, and read the agreement to learn what guest data comes back to you.
• 5. Food media, local best-of lists and city Reddit threads — the text you cannot edit. Semrush's AI Visibility Index (126 million prompts, January to April 2026; Semrush sells AI-visibility tracking) found ChatGPT cites 15.4 sources per answer, leaning on Reddit and Wikipedia, against Gemini's 3.3 (Semrush, via PPC Land). Do this: read your city's food threads for the words people use about your room, and answer them on your pages.
• 6. The delivery apps — a listing, a menu, and a commission of 15% to 30% (DoorDash: 15/25/30; Uber Eats: 20/25/30, both published rate cards). Never measured as an AI citation source for restaurants. Last because it is where the answer routes when the question is delivery, not because anyone showed it decides who gets named. Do this: know which tier you signed.

Seven questions a guest types before dinner. Run them yourself, this week, for nothing.

Any restaurant can measure this with no tool and no vendor. Open ChatGPT, Gemini, Perplexity and Claude, use a fresh chat per question, and paste the prompts below with your city in the brackets. Then record two columns, not one.

Column one is obvious: which restaurants get named, and whether yours is among them. Column two is what this article exists for — where does the answer send the guest to finish? A reservation platform, a delivery app, or your own phone number? Count the endings; nobody has published that ratio for anybody's restaurant. Then add one more prompt and watch the assistant try to act rather than answer: "Book me a table at [your restaurant] for Friday at 8pm."

One pass is one moment: ask twice and the names move. Run it anyway — a rough snapshot is still the first number your restaurant has ever had about itself.

• "Quiet place for a business lunch downtown in [city] where we can actually hear each other."
• "Gluten-free Italian in [city] that's actually safe for celiac."
• "Romantic table for two tonight in [city] — it's our anniversary."
• "Kid-friendly brunch with outdoor seating near [neighborhood]."
• "Late-night tacos in [city], somewhere still serving after 11pm."
• "Vegan options in [city] that aren't just a salad."
• "Birthday dinner for 10 with good cocktails in [city]."

What owners say when we show them this, including "if OpenTable fills my tables, why fight it?"

These come up in nearly every conversation with an owner, and more than one of them is right.

• "If OpenTable fills my tables, why fight it?" Then don't fight it — a full room on a Tuesday is a full room. The argument is narrower: know what each channel costs, and don't let one quietly become the only door. A guest routed to a booking platform is a guest whose reservation is created inside someone else's product, and what comes back to you is whatever your agreement says comes back. On the delivery side the same handoff has a printed price: 15% to 30% (DoorDash, Uber Eats). The channel was never free. It was billed where you weren't looking — and it became the only door because it is the only place your menu, your table and your hours exist in a form a machine can act on. Publish them yourself and the platform stops being the sole route to a guest the assistant had already decided to send you.
• "We're always full. We don't need this." Possibly true, and this is not an argument about Saturday. It is an argument about guests with no habit to fall back on — the anniversary, the client lunch, the family in town. They have to ask somebody, and 45% of US consumers now ask AI to recommend a local business, up from 6% (BrightLocal, n=1,002; sells local SEO and reputation tools).
• "Every number here comes from a company selling AI visibility." Almost every one, and each of them measures the thing it sells: Uberall sells AI-visibility software; SOCi sells local-marketing software; Yext sells listings management; Semrush sells AI-visibility tracking; BrightLocal sells local SEO tools. The exceptions are the two closest to your margin: DoorDash and Uber Eats published their own commission tiers. RecoSignal measures too — with the raw engine answers stored and the method published, so the number can be checked rather than taken on trust. Which is why this article ends in the free check above, not in an invoice.
• "Can I just come off the delivery apps?" A commercial decision, and we will not make it for you. But notice what nobody can tell you: there is no current open US source giving the share of restaurant orders that run through DoorDash, Uber Eats and Grubhub. Anyone quoting that percentage is quoting a number that does not exist. What is knowable is the rate card — 15/25/30 and 20/25/30 — and which tier you are on.
• "Isn't this just SEO with a new name?" Not cleanly — local-search strength does not transfer intact. In SOCi's 2026 index, only 45% of the top-20 brands in traditional local search reached the top of AI recommendations in retail (Search Engine Land; SOCi sells local-marketing software and has an interest in the gap it reports). See AI visibility versus local SEO.
• "Our menu changes every week. That's why it's a PDF." A real constraint, and it does not require dropping the PDF. Keep it, and add one plain-text page beside it with this week's dishes and the dietary and occasion words spelled out. Twenty minutes on a Monday, and it is the only version an engine can read.

Nobody has measured what share of restaurant orders the aggregators carry, which is why the only number worth having is your own

The hole in the middle of this subject deserves saying out loud. There is no published US figure for the share of restaurant orders flowing through delivery aggregators. There is no per-restaurant citation-share study of any kind. And the 83% that opens this article comes from a company selling the fix, in a release that does not fully disclose its sample (Uberall via BusinessWire).

Two things are solid. The commission tiers are published by the companies charging them: 15%, 25% and 30% at DoorDash (DoorDash); 20%, 25% and 30% at Uber Eats (Uber Eats). And OpenTable is OpenAI's first restaurant tech partner for Operator, the agent that books the table (OpenTable; OpenAI). Between them sits the argument of this page: the recommendation is winnable and costs nothing, and the transaction attached to it is routed by somebody else.

So the number to have is the one about your restaurant. RecoSignal's free AI Visibility Snapshot puts guest-style questions to ChatGPT, Gemini and Perplexity and reports how often your restaurant is named, which restaurants are named instead, and which sources those answers lean on. For what moves the odds, see how to improve your AI visibility; for one engine's mechanics, how to get recommended by ChatGPT.

It shifts probability, not certainty. Nobody can promise an assistant will name your restaurant, and there is no #1 inside ChatGPT to buy. What is in your hands is duller: a menu written in words, hours that are true, and a count of which channel the answer routes your guests through tonight.