How restaurants get cited by ChatGPT, Claude, and Gemini in 2026

A regular at a small ramen shop in Shibuya used to pull out his phone, open Tabelog, scroll through the noodle category, and tap the highest-rated option within walking distance. Today, he opens ChatGPT and types: "best ramen in Shibuya for under ¥2,000, open right now, English menu okay." He gets three answers in eight seconds. He picks one. He walks there. The restaurant he picked was never on the first page of Tabelog. It just happened to have the right structured data, an English menu page, and a clear `openingHoursSpecification` block that ChatGPT could read.

Multiply that interaction by the hundreds of millions of weekly active users on ChatGPT, Claude, Gemini, and Perplexity, and you have the beginning of a structural shift in how restaurants get discovered. We covered the macro picture in the middleman tax is dying. This piece is the restaurant operator's version of the same story — what is actually changing, what AI assistants actually read when they pick a restaurant, and what to do about it before your competitor figures it out.

The shift: AI is becoming the new "hey, where should we eat?"

Three things happened in a tight 18-month window that turned AI assistants into a real discovery channel for restaurants. August 2024: SearchGPT prototype launched, bringing live web grounding to ChatGPT for the first time. October 2024: ChatGPT search became generally available to all logged-in users. November 2024: Perplexity launched Shopping Hub, making its assistant a transactional surface, not just an answer engine. By early 2026, Google had rolled AI Overviews into the default search experience and Apple Intelligence had shipped on every iPhone sold since the 15 Pro.

The macro forecast matches the consumer behavior shift. McKinsey's October 2025 analysis projects up to $1 trillion of US B2C revenue could be orchestrated through AI agents by 2030. The qualitative version is more useful for a restaurant owner: assume that within five years, a meaningful slice of "where should we eat tonight?" stops happening on Tabelog, OpenTable, and Yelp, and starts happening inside whatever AI assistant the customer already has open.

Even today, the pattern is visible if you look. Open ChatGPT. Ask it: "good Italian restaurant in Brooklyn, vegetarian options, around $40 a head, books a table for tomorrow." You will get an answer with three to five named restaurants, complete with cited sources. Some of those citations come from Resy and OpenTable. But many come from the restaurants' own websites — the ones whose schema, menu pages, and content patterns make them legible to a language model.

The middleman tax restaurants pay today

Before getting into the AI playbook, it is worth being precise about what the current platform stack costs a typical restaurant. The fees are not subtle:

Delivery. Uber Eats charges 15-30% per delivery order depending on Lite, Plus, or Premium tier. DoorDash charges the same 15-30% range across Basic, Plus, and Premier. The headline rate is not the full story — both platforms also charge customer-side delivery fees and run paid-placement ads where restaurants bid against each other for a slot in the recommended carousel.

Reservations. OpenTable charges $1 per cover seated through OpenTable.com plus a monthly subscription ranging from $39 (Basic) through $249 (Core) to $449 (Pro). Resy charges flat monthly fees plus per-cover charges that vary by market and plan, and Yelp's reservation product is bundled into Yelp's advertising tiers. For a 60-cover restaurant doing two seatings five nights a week, OpenTable alone runs north of $25,000 a year before subscription fees.

Listings (Japan). Tabelog and Hot Pepper Gourmet's specific rates are not publicly disclosed — owners receive private quotes — with industry trade press citing premium plans in the ¥120,000 to ¥1.2 million per year range per restaurant, plus per-cover commissions on online reservations. The opacity is partly by design: a public price list would make the spread between top-tier and bottom-tier visibility too easy to compare.

Add it up. A typical neighborhood restaurant doing ¥30 million (~$200,000) a year in revenue, half delivery and half dine-in, can easily route 15-25% of revenue to a combination of Uber Eats / DoorDash / OpenTable / Tabelog before paying for food, rent, or wages. The platforms got rich. Customers got convenience. Restaurants got squeezed. None of these platforms control what ChatGPT, Claude, or Gemini cite. That is the door the AI shift cracks open.

What ChatGPT actually reads when answering "good Italian restaurant in Brooklyn"

There is no public ranking algorithm for AI assistant recommendations the way there is for Google search. But by reading the published bot documentation from OpenAI, Anthropic, and Perplexity, the Schema.org specification, and the Google structured-data guidelines, a clear list emerges of what these systems are looking for when they pick a restaurant to cite.

1. Schema.org Restaurant markup, with the right subtype

The single highest-leverage fix is correctly implemented Schema.org Restaurant markup. The Restaurant type extends `LocalBusiness` and adds restaurant-specific fields that AI assistants parse directly. The minimum viable markup includes `name`, full `address` (as a `PostalAddress`), `geo` coordinates at five or more decimal places, `telephone`, `priceRange`, `servesCuisine`, `acceptsReservations`, `openingHoursSpecification`, `menu` (as a URL or `Menu` node), and `sameAs` linking to your Wikidata Q-item, Google Business Profile, Instagram, and TripAdvisor. Google's structured-data guide for local businesses documents the exact required and recommended properties.

Use the most specific subtype. Schema.org has FastFoodRestaurant, BarOrPub, CafeOrCoffeeShop, IceCreamShop, Bakery, BreweryOrPub, Distillery, Winery, and several others as Restaurant subtypes. A ramen shop is a `Restaurant` with `servesCuisine: "Japanese ramen"`, not a `FastFoodRestaurant`. A kissaten is `CafeOrCoffeeShop`. Subtype precision is one of the cheapest disambiguation signals you can give an AI assistant.

2. robots.txt that explicitly allows AI search bots

AI-readiness fails silently if you forget this step. OpenAI documents three distinct user agents: `GPTBot` (used for training), `OAI-SearchBot` (used to surface citations in ChatGPT search answers), and `ChatGPT-User` (used when a logged-in user asks ChatGPT to fetch a specific page). Anthropic similarly documents `ClaudeBot`, `Claude-SearchBot`, and `Claude-User`. Perplexity documents `PerplexityBot` and `Perplexity-User`. If your robots.txt was generated three years ago, it almost certainly does not mention any of these — and if your CDN or WAF is on a default "block unknown bots" rule, you are invisible to AI search no matter how perfect your schema is. Add explicit `Allow:` rules for the search bots, and decide separately whether to block training bots.

3. Real text content — your menu cannot be a PDF or an image

AI crawlers do not document JavaScript execution, and they do not OCR images at scale. Google's own JavaScript SEO documentation confirms that not all bots run JavaScript, and that you should assume the worst case. Your menu, your story, your chef bios, and your address must be in raw HTML at server-render time. The single most common reason a beautiful restaurant website is invisible to ChatGPT is that the menu is a designer-built PDF that the bot never opens.

4. llms.txt — cheap insurance, low confirmed lift

Add an llms.txt file at the root of your domain following the Howard 2024 spec, with a one-paragraph restaurant description, a link to your menu URL, your hours, your reservations URL, and your address. No major AI lab has officially confirmed using llms.txt for citation, but several agentic developer tools and meta-search products consume it, and the file takes five minutes to write. Cheap insurance.

5. Entity disambiguation — sameAs to Wikidata, GBP, and Wikipedia

Language models triangulate entity identity across multiple sources to avoid confusing your ramen shop with another ramen shop of the same name in another city. Claim and verify your Google Business Profile (free, takes an hour, single biggest entity signal an AI assistant can triangulate against). Create a Wikidata item if one does not exist. Link them all in your Schema.org `sameAs` array. This is what separates a restaurant the AI is confident about from a restaurant the AI hedges on or skips entirely.

Apply the KDD 2024 GEO research to a restaurant page

The KDD 2024 GEO paper by Aggarwal et al., arXiv 2311.09735, tested seven content rewrites against Perplexity.ai and other AI engines to see which ones lifted citation rates. Three rewrites moved the needle hard. Two did not move it at all. One actively hurt. Translated into restaurant content, the findings are concrete:

• Direct quotes from named sources lifted citation by 41% — the largest single effect in the paper. For a restaurant: include direct quotes from your chef, your sommelier, your sourdough baker, or critics who reviewed you. "Our shoyu broth simmers for 14 hours with chicken bones from a single farm in Chiba," says chef Kenji Watanabe — that is the kind of sentence ChatGPT cites verbatim.
• Statistics with cited numbers lifted citation by 31% overall, +37% on Perplexity. For a restaurant: include the actual numbers. 14-hour broth. 47 dumplings folded per hour by hand. 1928 founding year. 6-table counter. The vagueness of "longstanding tradition" is exactly what AI assistants skip; specificity with provenance is what they pick up.
• Inline citations to authoritative sources lifted citation by 27% overall, and by +115% on currently low-ranked pages. For a restaurant: link to the Michelin Guide entry, the Tabelog page, the Eater article, the New York Times review, the Wikipedia page on the regional cuisine you serve. The +115% number is the most important finding in the paper for SMBs — it means the citation tactic actively democratizes who AI surfaces.
• Two rewrites did not lift citation at all: keyword density and authoritative tone alone.
• One rewrite hurt: keyword stuffing measured roughly 10% worse than baseline. Do not pack "best ramen Tokyo authentic Japanese noodles Shibuya" into your meta description; AI engines penalize it.

The paper's implication for a restaurant owner is straightforward: write the page in the voice of an editor profiling your restaurant, not a marketer selling it. Quote your chef. Cite your reviews. Use real numbers. Skip the SEO-tone adjectives.

What AI-readiness cannot fix

An honest counter to everything above. AI-readiness is a discovery layer; it is not a product fix. There are five problems no amount of Schema.org markup, llms.txt, or quoted-chef content will solve:

• Bad food. AI assistants increasingly weight aggregate review signals from Google, Yelp, Tabelog, and TripAdvisor when deciding which restaurant to cite. Real reviews from real customers about real food still dominate.
• A dirty venue. Health inspection records and review mentions of cleanliness propagate fast. AI engines pick them up.
• No real reviews. A restaurant with five Google reviews from the past two years is invisible to AI regardless of schema. Citation engines need corroboration.
• Genuinely terrible service. Recent negative reviews mentioning service kill citation lift. AI engines surface caveats ("reviews mention slow service") that no markup can hide.
• Pricing that does not match the experience. AI engines parse `priceRange` and cross-reference it against menu prices and review sentiment. Mismatches surface as caveats, not citations.

Get the offline product right first. Then make sure the AI can read it. The reverse order is a guaranteed waste of money.

The Monday playbook: 5 actions to do this week

If you are a restaurant owner and you have read this far, here is the concrete sequence. None of the first four cost any money. Start at the top.

• 1. Run the discoverability test. Open ChatGPT, Claude, and Perplexity in fresh sessions. Ask each: "best [your cuisine] in [your neighborhood]" three times. Record whether your restaurant appears, and which sources are cited when it does. This is your baseline. If you appear once in nine prompts, that is your starting score.
• 2. Audit your robots.txt. Add explicit `Allow: /` rules for `OAI-SearchBot`, `Claude-SearchBot`, and `PerplexityBot`. Decide separately whether to allow `GPTBot`, `ClaudeBot`, and `PerplexityBot` (training bots). The search bots are the ones that put you in answers; blocking them is the most common own-goal we see.
• 3. Claim and verify your Google Business Profile. Free, one hour at business.google.com. This is the highest-leverage entity signal an AI assistant can triangulate against. Your restaurant name, address, phone, hours, cuisine, price range, and photos all live here. Add the menu URL.
• 4. Get your menu out of PDFs and images and into raw HTML. If your menu is an image tag pointing at menu.jpg, or a link to menu.pdf, AI assistants cannot read it. The fix is a simple HTML page with the menu items, prices, and section headings.
• 5. Add Schema.org Restaurant markup to your homepage and menu page. Use the Schema.org Restaurant spec, the most specific subtype that fits, and the Google structured-data guide. Validate in Google's Rich Results Test before deploying.

After those five, the next layer of work — Wikidata creation, content rewrites against the GEO levers, internal linking architecture, llms.txt, sameAs graph completeness — is where most owners hit diminishing returns on DIY. That is the work our $47 audit scores for you, and the work our build packages ship as a one-time deliverable. No commission. No listing fee. No monthly subscription to anyone — including us — once the build is done.

The bet, restated for restaurants

AI shopping is not a magic wand. Most diners in 2026 still discover restaurants through Google, Instagram, friends, and a slow walk down a busy street. The AI-driven slice will take years to grow into the majority of inbound, and there is a real risk that AI search consolidates around one or two players who eventually charge for placement the way Google did. We covered that counter-thesis in the middleman tax is dying and we are not pretending it is solved.

But the work to be cited by ChatGPT, Claude, Gemini, and Perplexity is also the work to be trustworthy on the open web. Schema.org markup helps Google. Server-side rendering helps everyone. A clean menu page helps your human customers more than your AI ones. Allowing AI bots to crawl your site costs nothing if the AI shift fizzles, and it costs you everything if it doesn't and your competitor down the street did the work first. The middleman tax is dying. The question for every restaurant is whether you'll still be paying it when it's gone.