Nobody Actually Uses OpenTable Concierge — and That's Fine

I spent Tuesday night with a glass of something bracing and the OpenTable app open on my phone, typing the kinds of questions I imagined a diner typing into the new Concierge: what time do you open, can I see the menu, is there parking. I did this in a deliberately stupid way, the way you have to test a consumer AI feature if you want a read on it that isn’t a demo, and I did it against eleven restaurants across three cities, picked because I had eaten at all of them in the last year and I knew, roughly, what the right answers were. The interface is fine. The answers are mostly correct. The thing I cannot stop thinking about is that the most useful question I asked all night was what time do you open — a question that has been answerable on the OpenTable restaurant page in roughly 0.4 seconds since approximately 2003.

The contrarian thesis I want to put down before this gets pattern-matched as “diner-facing AI is the next platform shift”: the AI Concierge is not a product. It is infrastructure for data capture, and that is fine — it’s just not what the press release says. The July–September usage data OpenTable has now published makes the read unambiguous. Diners are asking the chat layer the same questions they would have typed into a search box or scrolled to find on a static page. The AI is doing the work of a slightly improved FAQ widget. What the AI is also doing, quietly, is generating structured intent data the platform did not have access to a quarter ago. That is the part of this launch that matters. That is the part the deck slides do not foreground.

This is a Desk Review, the format I use when a product has been in market long enough to read against its own data but not long enough to run a full ninety-day operator panel against. Concierge launched in July. We have roughly seven weeks of usage data and another month of post-Labor-Day signal. The rubric prints at the bottom in the usual format.

Methodology disclosure (the part Vibe Check requires up front)

A proper Vibe Check runs six operators against a named product on a structured weekly log for ninety days. Concierge is diner-facing, which means the operator panel framing has to be adapted — what an operator sees of Concierge is not the chat itself but the downstream effects in the reservation graph, the cancellation patterns, the “AI-assisted” diner tag, the new fields that started showing up in the OpenTable host dashboard mid-August. So I worked the desk side of the review instead. I read the OpenTable Dining Trends data page for the July 14–September 2 window the company published in early October, I went back to the July launch coverage in PYMNTS to anchor the launch claims against what the chat actually does today, and I read the Fox News piece on OpenTable’s diner-behavior tracking twice, the second time with a notebook, because it is the most honest writing I have seen this year about what the platform actually does with the dossier it keeps on each diner. The forthcoming May piece on OpenTable’s place in the Booking Holdings AI strategy reads against the same Q1 transcript I keep returning to. Treat that piece as the corporate-strategy frame. This one is the product read.

I also typed into Concierge for roughly two hours across the eleven restaurants. The sample is small. The pattern was clean enough that I am comfortable filing.

What the July–September data actually says

OpenTable published the first usage window for Concierge on October 7. The numbers are the part of this story that everyone trying to write a tidy AI-for-restaurants narrative is going to have to wrestle with. Forty-six percent of diner questions in the window were about opening hours. Thirty-nine percent were about the menu. Twenty-two percent were about the address. Those three categories alone cover the great majority of inbound queries. The long tail — wine list questions, dietary accommodations, parking, dress code, “do you have a back room for ten people” — is real but it is not where the volume lives. The volume lives in the three questions a static restaurant landing page already answers.

Two reads of that data, both of which I think are correct.

The first read is the one OpenTable’s product team will give you, and it is not wrong. People ask basic questions. A consumer AI surface that answers basic questions instantly, in natural language, on a phone, while the diner is already inside an app that converts to a reservation, is a meaningful UX improvement over the diner having to scroll, tap into a separate tab, parse a poorly-formatted hours block, and then come back to make the booking. The conversion lift on the booking flow — if there is one, and the company has not disclosed numbers — is the metric that will eventually justify the build. I am not arguing that Concierge is bad. I am arguing that it is doing something much smaller than the launch material implies.

The second read is the one the operator and the strategist both have to make. If 46 percent of the questions are what time do you open, the AI is not solving a hard problem. It is solving a discovery problem. The model is doing trivial retrieval against structured data the platform has had for fifteen years. The cleverness is in the wrapper, not the answer. A well-tuned natural-language search box from 2018 would handle most of this volume. The chat layer adds polish; it does not add capability. And — this is the part the AI press is not writing about — the chat layer also tells OpenTable something the static page never could: who is asking, what they are asking, when, and in what sequence. That is the data point. That is what the launch is actually for.

The data capture is the product

I want to dwell on this for a minute, because it is the part of Concierge that the launch coverage almost completely missed.

Before Concierge, OpenTable knew when you searched a restaurant, when you booked, when you showed up or didn’t, what you ordered if the venue is on an integrated POS, who you dined with if they were also OpenTable users, and a slow-accumulating profile of cuisines, neighborhoods, price points and times you favor. That dossier — the Fox News piece is sharp on this, and worth reading if you have not — is the asset Booking Holdings paid $2.6 billion for in 2014 and the asset every subsequent AI investment is leveraging. The dossier was always behavioral. It captured what you did. It did not capture what you intended.

Concierge captures intent.

When a diner types do you have anything gluten-free into the chat, the platform now has a structured signal — at the diner-ID level, the restaurant-ID level, and the timestamp level — that the diner cares about gluten in a way they may not have cared about it on any prior visit. When a diner types can I bring my dog, the platform learns the diner has a dog. When a diner types is there a corkage fee, the platform learns the diner is the kind of diner who brings their own bottle. None of this is in the booking record. None of this is in the post-meal review form. All of it is now in the Concierge transcript log, and all of it is presumably joining the diner dossier at the level of structured tags.

This is the part of the launch the AI-ethics conversation has not caught up to yet, and the Fox News piece is the only mainstream writeup I have seen that even gestures at it. The article quotes OpenTable on the volume and specificity of the data the platform retains per diner — order history, allergy notes, special-occasion tags, behavioral patterns across multiple restaurants — and the framing is the one operators ought to be reading. Concierge is the latest, and most efficient, intake funnel for that dossier. The chat is the wrapper. The dossier is the product.

I am not saying this is bad. I am saying it is the actual product. The launch material says Concierge is “an AI assistant for diners.” That is true in the same way that a free email service is an email service. It is true and incomplete.

The “AI-assisted” diner tag, in the wild

Mid-August, an operator I have been talking to for a year — a mid-volume New American spot in San Francisco that uses OpenTable for the floor and Resy for waitlist — started seeing a new column in his host dashboard. The column was labeled AI-assisted and it appeared, intermittently, against guest reservations. He could not find documentation for it inside the dashboard. He could find it referenced obliquely in a release-notes email from late July that I subsequently went back and read. The tag flags reservations where Concierge interactions produced what OpenTable internally calls signals worth surfacing to the venue.

In the period from mid-August through late September, the operator saw the tag attach to two categories of reservations with enough consistency that he started screenshotting them. The first was reservations that had been preceded by Concierge questions about drinks — specifically, questions about cocktails, wine pairings, or non-alcoholic options. The second was reservations the host team subsequently flagged as last-minute cancellations or no-shows. The platform was, in effect, telling the venue this booking came with conversational signal you might want to act on — the drinks pattern as a service-prep cue, the cancellation pattern as a risk cue.

That is a useful piece of operational data. It is also a meaningful escalation of what the platform does with the chat transcript. The Concierge log is no longer just diner-facing UX. It is feeding the host dashboard. That is infrastructure, in the precise sense — a piece of platform plumbing that produces value at a layer the diner never sees and the host has not yet been trained to read.

I asked the operator what he does with the AI-assisted tag. He said: nothing, for now. He does not trust it enough yet to change his cover counts or his prep, and he does not have a workflow that ingests it. He looks at it. That is the honest state of the product on the operator side seven weeks in. The platform is generating signal faster than venues can absorb it. That is a familiar pattern in restaurant tech — Toast and Square both shipped analytics layers years ahead of operator capacity to use them — and it is not, by itself, a problem. It is the early shape of a feature that will matter in a year and does not, today.

What Concierge gets right (and what the demo papers over)

I want to be fair to the product, because the parts that work are real.

The natural-language interface for basic restaurant questions is genuinely an improvement on the prior surface. The eleven-restaurant test I ran turned up correct answers to hours, menu, address and parking in all but one case (a restaurant whose listed hours on the OpenTable page were out of sync with the venue’s actual Tuesday hours; Concierge dutifully repeated the wrong data). The dietary-restriction queries returned plausible responses in most cases and a flat “please ask the venue directly” in two cases, which is the right answer when the data does not exist. The wine-list queries were the weakest — Concierge has access to whatever wine inventory the restaurant has published into OpenTable, which is rarely the full list, so the answers are partial in a way the diner has no easy way to detect. I would rate the surface a B on accuracy and a B+ on UX. It is a competent consumer AI feature.

The part the demo papers over is the part the usage data exposes. Concierge is not, in any meaningful sense, answering hard questions. It is answering easy questions in a nicer wrapper. The demos always show the dietary-restriction case or the special-occasion case, because those are the cases where the chat layer is doing something a static page cannot. The usage data says those cases are the long tail. The head of the distribution is hours and menu, and the head of the distribution is what the platform is built for.

This is the second time in a year I have written some version of this paragraph about a consumer-facing restaurant AI launch. The pattern is consistent. The demos showcase the long tail. The data shows the head. The platform value is captured in the data layer, not the model layer.

The competitive read

OpenTable is doing this because Resy is not yet, because Yelp is unlikely to ship anything comparable, and because the parent company — Booking Holdings — has a corporate-level mandate to put a Concierge layer across every reservation surface it owns. The forthcoming May piece on OpenTable’s place in the Booking Holdings AI strategy will work through the Q1 transcript in more detail, but the read is straightforward: Concierge is the restaurant-shaped version of a chat-and-book pattern Booking is running across hotels, flights and Priceline. The diner-side AI is not the moat. The integration with the reservation flow is the moat. OpenTable already owns that flow.

That means the competitive question is not will Resy ship a Concierge clone, because they probably will, but will any reservation platform without OpenTable’s coverage produce enough Concierge interactions to build the same intent dossier. The answer is almost certainly no. Coverage breeds data. Data trains better Concierge. Better Concierge produces more interactions. The loop is the moat, and the loop is six months old, and OpenTable is the only player running it at scale today.

For Toast and Square, both of whom have been building diner-facing surfaces of their own — Toast through Online Ordering, Square through the rebuilt Square Go consumer app — the read is that the AI Concierge layer is going to be a category-wide expectation in 2026 and that none of them are positioned to build it without the booking graph OpenTable has. They will build something. It will not be as effective. The Vol. 2 voice-ordering layer Square announced at the start of October is the inverse pattern — voice on the venue side, capturing the same kind of conversational intent on the inbound call surface that Concierge is capturing on the diner-app surface. Different funnels, same data play.

What operators should do today

Three things.

First, look at the AI-assisted tag in the host dashboard, even if you do not yet have a workflow that uses it. The tag is going to become more granular over the next two quarters. Operators who have spent the time looking at the early version will be faster at reading the mature version. The cost of the look is five minutes a week.

Second, do not bother building diner-facing AI of your own. The platform is doing it. The venue-level chat widget that a half-dozen vendors are pitching as your restaurant’s own AI assistant is going to be dominated by the OpenTable surface on volume, because diners are inside the OpenTable app already and they are going to ask Concierge before they ever land on your website. The venue-level AI build is a 2024 conversation. The 2025 conversation is whether you are surfacing the right data into OpenTable for Concierge to repeat correctly. Your hours, your menu, your dietary tags, your wine list, your parking field — those are now AI training data for a diner-facing surface you do not control. Get them right.

Third, watch the cancellation-risk version of the AI-assisted tag. If the tag matures into a usable last-minute-cancellation predictor — and the early operator signal suggests it might — that is the feature that pays for the rest of the build, because no-shows and last-minute cancels are the single largest unrecovered revenue line in mid-volume hospitality. A reliable predictor would change deposit policy, overbooking ratios, and floor-management practice across the industry. That is the feature to grade Concierge on in twelve months.

Vibe Check rubric: Concierge (Desk Review, Oct 15)

Axis	Mark	Note
Product clarity	B−	The “AI assistant for diners” framing oversells the model layer and undersells the data layer.
Usage signal	B	The July–Sept window is clean enough to read; the distribution is unflattering to the AI framing.
Operator value, today	C+	The AI-assisted tag is the most interesting line in the dashboard and the least documented.
Operator value, twelve months	B+	Cancellation prediction, drinks-prep signal, dietary surfacing — the data play matures into operator utility.
Platform moat	A−	Coverage × intent data × reservation flow. OpenTable is the only player running the full loop.
Recommendation	Hold	Do not build your own diner-side AI. Get your OpenTable listing data clean. Watch the host dashboard.

The vibe check is that Concierge is a real product doing a small job competently while quietly building infrastructure for a much larger one. The launch coverage graded the wrong axis. The data play is the play.

Nobody is actually using Concierge. And that is fine. That was never what it was for.

— Sofia leads Vibe Check vendor reviews for TableTransfers. Tips: [email protected].