ChatGPT Books Your Table Now: Inside the OpenTable, Resy, and Yelp Integrations

For about a year, the standard demo of an "AI agent" booking dinner ended the same way: the model found the restaurant, showed you the times, and then politely gave up. In a hands-on test published April 29, 2026, PCWorld watched Claude surface a live Resy availability grid for The Smith and then say, verbatim, "I'm not able to complete the booking on your behalf." ChatGPT's agent mode fared no better — five minutes of slow clicking through OpenTable's web UI, a party-size dropdown it got wrong ("It seems I may have mistakenly selected '2 people' instead of '3 people'"), and a final handoff: "Please take over to enter those details."

On August 10, 2026, that gap closed. Yelp announced that ChatGPT users can now access Yelp Reservations and Waitlist to book a table or join a queue at thousands of restaurants across the US and Canada without leaving the chat. Resy, owned by American Express, launched Resy Reservations in ChatGPT for US restaurants the same week. OpenTable powers recommendations globally. This is not a better restaurant finder. It is the first mass-market instance of an assistant completing a real-world transaction inside the conversation — and it is the concrete thing the agentic commerce stack has been building toward while everyone argued about why AI agents haven't gone mainstream.

TL;DR — what people are asking

How does booking inside ChatGPT actually work?

Start with what is publicly confirmed and what is inference, because a lot of coverage blurs the two.

Confirmed. Yelp's announcement (authored by Chad Richard, SVP of Business and Corporate Development) describes an expansion of a content partnership struck a few weeks earlier in late July 2026, which had already brought Yelp reviews, ratings, photos, and business attributes into ChatGPT answers. The new layer adds a booking form rendered in the chat. Resy's version surfaces real-time reservation availability directly beneath each venue in the recommendation list, with in-conversation refinement by time, party size, cuisine, or price.

Not confirmed. None of the three partners published a protocol name, an endpoint spec, or a schema.

What the shape tells you. An interactive card that renders inside the conversation, holds its own state across turns, refreshes live inventory, and submits a form is not something a plain text tool-call produces. It is exactly the shape of OpenAI's Apps SDK, which is built on the Model Context Protocol. In that architecture, a partner runs an MCP server that declares tools — think search_availability(location, party_size, window) and create_reservation(venue_id, slot_id, guest) — alongside an optional web UI rendered in a sandboxed iframe, in one of three display modes (inline, picture-in-picture, fullscreen). The Apps SDK launched with Booking.com, Canva, Coursera, Expedia, Figma, Spotify, and Zillow as pilot partners; a restaurant reservation platform is the most natural possible fifth verb for that surface.

The tell that this is not a private one-off is the parallel deployment. American Express also connected Resy to Claude, where diners browse live tables and reserve inside the conversation — and Anthropic's own connector ecosystem is MCP end to end (see our Claude connectors and MCP servers guide). One reservation platform shipping the same capability into two different assistants in the same window is what a protocol looks like from the outside. A bespoke integration does not replicate that cheaply. It is the same convergence we tracked when agent plugins standardized across OpenAI, AWS, Cursor, GitHub, and VS Code.

Three layers, and which one broke

The third row is the whole story. Everything else was already there.

Recommend vs. transact: why this is a bigger jump than it looks

A recommender is allowed to be wrong. If ChatGPT suggests a mediocre Thai place, you shrug and pick another. The cost of a bad output is one wasted paragraph.

A transactor is not allowed to be wrong in the same way, because its output has side effects in someone else's system. A hallucinated recommendation is annoying; a hallucinated reservation consumes a real table at a real restaurant on a real Friday night, and someone eats the loss. Crossing that line changes four things simultaneously:

1. Idempotency becomes mandatory. A retried recommendation is free. A retried create_reservation is a double booking. Every write path an assistant touches needs an idempotency key and a dedup window, because the model will retry on a timeout.
2. Confirmation becomes a contract, not a sentence. The user needs a durable artifact — a confirmation code that exists in the platform's database — not a fluent model utterance saying "you're all set." Research on task-oriented dialogue agents documents exactly this failure: when a backend returns empty results or errors out, models still produce fluent but unsupported responses, including invented confirmations.
3. Authentication has to happen somewhere. Party size and a phone number are PII moving from an OpenAI surface into a partner system. Someone is the data controller, and the answer differs per partner.
4. Reversal needs a path. Cancellation, modification, and no-show policy all have to live somewhere the user can reach after the conversation is closed. Yelp answered this the conservative way: changes are made through Yelp, not through ChatGPT.

That last decision is the most informative one in the entire launch, and it leads directly to the ownership question.

Who owns the customer and the booking data now?

Three parties, three different assets, and only one of them is new.

The strategic logic for Yelp is blunt and worth stating plainly: Yelp depends on Google for more than 70% of its US web traffic. Every booking that originates inside ChatGPT is a booking that did not require a Google SERP. Yelp already syndicates its content into Apple Maps, Alexa+, Bing, DuckDuckGo, and Yahoo, and launched Yelp Assistant in April 2026. The ChatGPT integration is the same de-risking play, executed at the point where the money actually changes hands. Yelp's own framing is that restaurants on Guest Manager gain access to "the largest restaurant consumer network in the US," helping them "fill more tables and reach more guests exactly when discovery happens."

PYMNTS put the underlying market shift more sharply: "if the AI answer is the last stop before a purchase, whoever controls that answer controls the sale." The supporting numbers are not marginal — 48% of ecommerce shoppers used AI during their most recent purchase, AI-referred traffic to US retail sites rose 393% year over year in Q1 2026, and ChatGPT's share of product research went from 2% to 30% in two years. Yelp itself sits on 330 million local business reviews, and claims a 3.4x citation lead over competitors as the most-cited home services platform across major AI systems — a self-reported figure, but a coherent one given the syndication footprint.

Note what the restaurant loses: the pre-visit brand impression. No homepage, no photography direction, no menu design, no carefully chosen hero image. The venue is reduced to a row in a comparison list with a price band, a cuisine tag, and four time chips. That is the tradeoff nobody in the announcements mentions.

What this means for local SEO and GEO

Here is the counterintuitive part for anyone who does local search professionally: there is no ChatGPT listing to optimize. There is no dashboard, no submission form, no schema tag that makes you bookable. You inherit the channel by being a live, correct record on a platform the assistant already trusts.

That collapses local visibility into a supply-chain problem. Practical priorities, in order:

1. Be on a bookable platform with live inventory. A restaurant with a phone number and no Yelp Guest Manager, Resy, or OpenTable presence is invisible at the transaction layer no matter how good its website is. Discovery without a bookable slot loses to a competitor with one.
2. Fix the structured attributes, not the prose. Cuisine, price band, hours, party-size limits, neighborhood, and dietary tags are the fields the model filters on when a user says "under $50, vegetarian-friendly, 7:30 for four." Missing attributes fail the filter silently — you are not ranked lower, you are excluded.
3. Keep availability honest. A venue that shows slots it cannot honor gets a bad first booking, and the correction loop runs through the platform's reliability signals, not through your marketing.
4. Review recency still matters, but as retrieval fuel. The model reads review text to justify a recommendation. Specific, recent reviews mentioning concrete dishes and occasions give it something quotable; a wall of five-star "great!" gives it nothing.
5. Do the ordinary GEO work on your own site anyway. Assistants triangulate. Our GEO guide and the SEO/GEO agent skill breakdown cover the citation-and-statistics tactics that get any entity described accurately.

The broader pattern: Google is testing agentic reservations in AI Mode via Labs, but its current experiment directs users to the restaurant's own booking page rather than completing the booking. OpenAI shipped the completion. Whichever approach wins determines whether the assistant is a referrer or a merchant of record — and those are very different businesses.

What breaks: failure modes and trust questions

Nobody publishes the error rate for these systems, so the honest version is to name the failure classes and check which ones the design actually mitigates.

The legal question sits underneath all of them. In Moffatt v. Air Canada, a Canadian tribunal held the airline responsible for incorrect information its own chatbot gave a customer — the airline's argument that the bot was a separate entity was rejected. That precedent is about a business's own bot, not a third-party assistant, but it establishes the direction of travel: "the bot said it" is not a defense. When a booking originates in ChatGPT, is fulfilled by Resy, and is honored (or not) by a restaurant, the chain of responsibility has three links and no published contract between the ends of it.

This is the same trust gap that made OpenClaw cancelling a stranger's gym booking a news story rather than a footnote. Write access to real-world systems is where agent reliability stops being an engineering metric and starts being someone's ruined evening.

For builders: what surface do you need to be bookable by an assistant?

This is the part that generalizes beyond restaurants. The pattern is discovery → real-time availability → in-chat transaction, and it applies identically to clinics, salons, tennis courts, tutors, repair shops, and any inventory that is time-sliced and perishable. If you run one of those, here is the surface you need to expose.

1. A structured availability read, not a rendered page.

GET /availability?entity_id=...&party_size=4&window=2026-08-14T18:00/22:00
→ { slots: [ { slot_id, starts_at, duration_min, holds_until } ] }

The critical field is holds_until. Conversations take minutes, not milliseconds. A slot the model surfaces at turn 3 must still be claimable at turn 7, or you have built a system that reliably shows users things they cannot have. Either return a soft-hold TTL or accept that your confirmation step will fail intermittently for reasons the user will blame on the assistant.

2. An idempotent write that returns a durable ID.

POST /reservations
Idempotency-Key: <uuid>
→ { confirmation_id, status, manage_url, cancel_by }

confirmation_id and manage_url are what make the interaction real rather than conversational. Return them, and the assistant has something true to say. Omit them, and the model will fill the gap with a plausible sentence — which is precisely the invented-confirmation failure the dialogue-agent literature documents.

3. Canonical entity identity.

Assistants disambiguate before they transact. Give them a stable entity_id, a canonical name, a geocoded address, and cross-references to whatever public identifiers you already have. "The Smith" is four restaurants; the model needs to know which one it is holding a table at.

4. A machine-readable policy surface.

Cancellation window, deposit amount, no-show fee, maximum party size, minimum notice. Put these in the tool response, not in a PDF on your website. An agent that cannot read your policy will either refuse to book or agree to terms you did not offer.

5. Pick your protocol posture deliberately.

If you expose these as MCP tools, you are portable across ChatGPT, Claude, and every other MCP host by construction — which is exactly how Resy ended up in two assistants at once. If you build a private partner integration, you get one channel and a renegotiation every time a new assistant matters. Our build your first MCP server walkthrough covers the mechanics, and the MCP security guide covers what you are exposing when you hand write access to a model.

6. Decide where money enters.

Reservations are the easy case — they move a record, not a payment. Deposits, prepaid tickets, and ticketed tasting menus need actual rails, which is where the Agentic Commerce Protocol (OpenAI and Stripe), Cloudflare Wallets, and Mastercard Agent Pay for Machines enter the picture. Restaurant booking is the deliberately low-stakes on-ramp: high volume, high intent, and the worst case is an empty table rather than a chargeback.

The bottom line

The August 10, 2026 launch is not a feature; it is a proof that the last mile closed. Four months earlier, the state of the art was Claude saying it could not complete a booking and ChatGPT's agent mode misreading a dropdown after five minutes of clicking. Now three of the largest reservation networks — OpenTable globally, Resy in the US, Yelp across the US and Canada — commit real bookings inside the conversation, with post-booking control deliberately retained on the platform side.

For consumers, the win is small and real: one fewer app switch. For reservation platforms, it is a hedge against Google dependence at the exact moment of conversion. For restaurants, it is demand at the decision point in exchange for the pre-visit brand impression. And for builders, it is a template with a short spec: expose live availability with a hold window, an idempotent write returning a durable confirmation, canonical identity, and readable policy — over a protocol that more than one assistant speaks. Everything after that is which vertical goes next.

Related on explainx.ai

• What is MCP? Model Context Protocol complete guide — the standard the Apps SDK and Claude connectors are both built on
• Build your first MCP server, step by step — how to expose the availability-and-booking tools described above
• How to use Claude connectors and MCP servers — the other assistant Resy shipped into
• MCP security guide 2026 — what you are handing over when an assistant gets write access
• Stripe Directory: the infrastructure layer that makes AI agent commerce real — the discovery-and-identity half of agentic commerce
• Cloudflare Wallets: programmable payments for AI agents — where money enters once bookings need deposits
• Mastercard Agent Pay for Machines (AP4M) explained — card-network rails for agent-initiated purchases
• Cloudflare Monetization Gateway: x402 for APIs and MCP tools — per-call pricing for the tools agents consume
• Why AI agents haven't gone mainstream (yet) — the adoption argument this launch is a direct test of
• Agent plugins: the OpenAI standard across AWS, Cursor, GitHub, and VS Code — the same protocol convergence, on the developer-tools side
• OpenClaw cancelled a stranger's gym booking — what write access to real bookings looks like when it goes wrong
• What is SEO/GEO? Generative engine optimization in 2026 — the visibility playbook behind the local-search section
• SEO/GEO agent skill for AI search — the tactical version, as an agent skill
• GPT-5.6 Sol now runs all of ChatGPT — the model tier doing the reasoning behind these bookings
• AI for travel planning: complete guide — the adjacent vertical where the same pattern is furthest along
• What are AI agents? A plain-English beginner's guide — the recommend-versus-transact distinction, from first principles

Official sources: Yelp Brings Reservations and Waitlist to ChatGPT — Yelp Official Blog, Chad Richard, August 10, 2026 · Yelp brings restaurant reservations and waitlists to ChatGPT — Search Engine Land · Yelp Lets Users Book Tables From Within ChatGPT — PYMNTS · ChatGPT users can now book tables and join restaurant waitlists through Yelp — 9to5Mac · ChatGPT and Claude tried to book my dinner. It got clunky fast — PCWorld, April 29, 2026

Integration details reflect announcements dated August 10, 2026 and press coverage published the same week. Coverage areas (OpenTable global, Resy US, Yelp US and Canada), rollout across ChatGPT plans, and participating venue counts are as stated by the partners at launch and may change. The underlying protocol is inferred from observable behavior and the Apps SDK architecture, not confirmed by OpenAI, Yelp, Resy, or OpenTable — verify against official developer documentation before building against it.