.jpg)
🏨 Google Will Now Take The Booking Itself
.jpg)
Agentic hotel booking went live in AI Mode this week. Ten launch partners, payment inside the chat, and almost none of the journey data coming back down the line.
Google has spent nine months teasing agentic travel booking. As of Thursday it is real.
Travellers in the US can now search, compare, check cancellation terms and pay for a hotel room without ever leaving the AI Mode chat window. Payment runs through Google Pay. The hotel or OTA still charges the card, sends the confirmation and handles support, so Google can insist it isn't an OTA - it's a transaction facilitator.
Launch partners are not small: Booking.com, Expedia, Hilton, Marriott, IHG, Choice, Wyndham, Priceline, Hotels.com and Trip.com.
Here is the part that should worry every marketer reading this.
🎟️ Partners get the minimum and nothing more - name, payment, occupancy, property. That's it.
📤 Chat context is withheld - the questions asked, the options rejected, the reasons given. Google keeps all of it, even in aggregate.
🎯 The path to conversion goes dark - you get a booking with no story attached to it.
US only for now - UK operators will have to wait a while.
✈️ Flights + reservations are still outside AI Mode - hotels are the test case for whether people will actually transact in a chat.
This isn't happening in isolation. Radisson put over 1,000 hotels inside ChatGPT in late July with Accenture. IHG got there first in June with 7,000+, and has since added conversational search to IHG.com and the One Rewards app. Drury announced one in July. Marriott built Ask Bonvoy on its own channels. Amazon's Alexa+ is closing travel bookings in-chat with Priceline, Viator and Virgin Atlantic.
Two different bets are being placed. Owned channels buy you control and data. Third-party assistants buy you reach and cost you both.
There is no equivalent for restaurant bookings yet. There will be, and the qualifying criteria are already visible: structured data, live availability exposed through an API, and a booking system a machine can actually transact against. The venues that get surfaced will be the ones a model can read, not the ones with the best-written landing page.
Being machine-actionable is starting to matter more than being well-optimised.
---
🔍 Your Search Console Is Full Of People Saying "Yes"
.jpg)
AI Mode follow-ups are being logged as queries in GSC. They look like nonsense data, and the report built to isolate AI performance can't show them to you.
Ask a follow-up question inside AI Mode and you have, as far as Google is concerned, performed a brand new query. All the impression, position and click data attaches to that new query. Say "yes" to an AI Mode prompt and "yes" enters your query data.
The awkward bit is where that data lives..
📊 AI Mode data sits inside your Web totals - bundled into the main performance report, unlabelled.
🚫 The Generative AI report has no query dimension - Pages, Countries, Dates, Devices and impressions, and that's the lot.
❌ No clicks or CTR in there either - you can see you appeared, not whether anyone came.
🔎 The usual cleanup trick fails - most teams filter for ten-plus-word prompts to spot AI queries; a one-word affirmation sails straight past.
📉 Nobody has measured how common this is - no published rate, no benchmark, no way to size it yet.
So your query data is now being distorted at both ends. Long conversational prompts at one extreme, one-word fragments at the other, and the report designed to separate AI visibility from everything else is structurally incapable of showing you either.
For hospitality clients the practical hit is on branded and near-branded reporting. If your Search Console query table is feeding a monthly deck, it is quietly getting noisier every month and nobody has flagged it.
Worth doing now: pull your last ninety days, sort ascending by query length, and see what's sitting at the bottom. If there are affirmations in there, your baseline has already moved.
Measure the trend, not the table.
---
🍽️ SevenRooms Can Now Search Guests Who Have Never Been To Your Restaurant
.jpg)
Global Guest Search opens up the DoorDash and Deliveroo customer base to partner venues. The commercial logic is obvious. The data questions are not.
Regular readers will have followed this thread. Echo 011 covered DoorDash buying both SevenRooms and Deliveroo. Echo 012 covered Channel Connect. This is the next move, and it's the biggest one.
On 21 July, SevenRooms launched Global Guest Search, letting partner restaurants search what the company describes as hundreds of millions of DoorDash and Deliveroo consumers.
The pitch is guest recognition: your host can greet a first-time diner already knowing their usual order. Alongside it came Client Insights, which compiles spending patterns, reservation history, guest notes and staff feedback into a profile without anyone doing manual research.
Set against Deliveroo Reservations launching in London in May with Dishoom, Barrafina, Kricket and Kolae on board, the shape of the strategy is clear. DoorDash is building a single guest identity across delivery and dining, and renting it back to operators.
The commercial case is strong.
The unanswered questions are the interesting part.
❓ What actually crosses? A name, a dietary note, an order history, a spend band? The announcement covers the feature, not the fields.
❓ Which consumers are eligible? Hundreds of millions is a big number, and it plainly includes people who have never dealt with your venue.
❓ Which markets is this live in? UK and EU consent rules are not US consent rules.
❓ Who is the controller? SevenRooms sits as a processor for the venue, but a restaurant receiving personal data about someone who has never been a customer starts to look like a controller with its own obligations.
❓ What is the lawful basis? Deliveroo's UK customer policy describes sharing within the DoorDash group where needed to fulfil an order. Recognising a guest at a different restaurant is a different purpose
None of this makes the feature unlawful, and there are perfectly good answers that would settle it. They just haven't been published yet.
What matters for operators is where the liability sits. If a guest asks how you knew their usual order, the ICO complaint lands on the restaurant, not the platform. Ask your provider for the field list, the eligibility criteria, the markets, and the controller position in writing before you switch it on.
Ask the questions before the feature does something clever on your behalf.
---
🎚️ Google Is Testing A Volume Knob For PMax Channels
.jpg)
A new Channels setting lets you tell the algorithm which surfaces you actually value. It's in alpha, and it's not the budget control everyone thinks it is.
The setting shows sliders for Discover, Gmail, the Display Network, Search, Search partners, Maps and YouTube. Move one up and you tell Google you value conversions from that channel more. Move it down and you say the opposite.
Read the mechanics carefully, because this is not what most people assume.
⚖️ It adjusts CPA tolerance, not budget - a positive adjustment relaxes the CPA the system will accept on that channel, a negative one tightens it.
🚫 You still can't split spend by channel - nobody is getting "30% to YouTube".
🧪 It's labelled Alpha - that's earlier than beta, so the "contact your rep" advice doing the rounds is optimistic.
📊 It only works if your channel reporting is trustworthy - garbage in, expensive garbage out.
The obvious use case is real. If PMax has parked a chunk of budget in a channel that reads badly and you've been unable to do anything about it, this is the lever.
The trap is attribution. Channel-level performance and channel-level value are not the same thing. A guest discovers you on YouTube, thinks about it for a fortnight, then converts on search. Tighten YouTube because its directly attributed CPA looks poor and you may quietly cut the demand that Search was harvesting.
For hospitality brands running awareness-heavy video, that risk is not theoretical. If you get access, move one channel at a time and hold it for a fortnight. Then look at total campaign conversions, not the channel you touched.
A control you don't understand is just a faster way to break something.
.jpg)
.jpg)