New study: Meta's Muse and Instinct travel bookings

In August, our study showed that travelers hand AI their trip research, but stop during the transaction to handle payment themselves. When attempting to pass booking details to the travel sites, 75% of handoffs from the LLM lost their details or failed, and people blamed the travel site.

Two months later: new consumer-friendly AI agents have launched that can transact, and we believe that it will change the game.

In this issue:

→ Meta’s Muse and Instinct can now take a booking from a text message to a confirmation email, and we’re in the field measuring that experience. Sign up for the new report below.

→ Drew Fortner of Alaska Shore Tours wired an AI agent into his reservation system, and the dashboard keeps landing near a 9,000% return.

→ Your ChatGPT traffic didn’t collapse on October 9th - they just accidentally stopped tagging it.

→ When Skift asked Muse for a hotel, it opened a browser and picked an OTA on its own. The airlines in the same search had a signed deal.

→ Google is handing back channel controls, with a default change in February that costs money if you ignore it.

Let’s get into it.

‍

NEW RESEARCH

Our Next Study Is in Field: Agentic Travel Booking

Our first two studies watched travelers search and plan with AI. With the launch of Meta Muse and Instinct, consumers are starting to transact directly with AI agents. Noah Shinn, the Founder of Instinct, reported nearly 50% of their first $1B transactions inside Instinct were all travel-related.

We believe this has significant implications to travel discovery, and our newest report studies the experience.

Meta’s Muse and Instinct now take a trip from a text message to a confirmation email. The traveler types one sentence, the agent picks the room and the rate, and stops on one button: "Confirm payment?"

We explored what happens inside that, and are measuring it in two ways:

What the agents do. The same booking scenarios, run through Muse and Instinct again and again. Where each one routes the booking, which merchant it lands on, where it fails.

How people use them. A survey of current users on when they take over from the agent and why, and live sessions watching the exact moment they decide to.

One shows the behavior, the other shows how common it is and why.

Together: when a traveler hands the booking to an agent, what does the agent do with it, and when does the traveler step back in?

The full study publishes on November 10th, with a live webinar the same day going through all the findings.

📅 Tuesday, November 10

🕚 10:30 AM Central US

Save Your Seat

‍

TRAVEL MARKETING PODCAST

How A Shore Excursion Marketplace Put AI To Work On Bookings

The Gist

→ Alaska Shore Tours wired an AI agent into their reservation system, so it can check a guest’s cruise schedule mid-chat and hand back a ready-to-book proposal.

→ They measure AI spend against minutes saved, and the dashboard keeps landing near a 9,000% return.

→ 80% of their customers book within 24 hours of finding them, which changes what top-of-funnel spend is for. spend is for.

Brennen sits down with Drew Fortner, founder and CEO of Alaska Shore Tours, a shore excursion marketplace that has planned tours for over 125,000 cruise guests and runs seven regional booking sites across Alaska. Drew started as a tour guide in Colorado before moving into sales and marketing, which is where he saw what diversified revenue could do for operators.

His team has since wired an AI agent straight into their reservation system. In the middle of a web chat it can check a guest’s cruise schedule and drop a ready-to-book proposal, a build that five years ago would have cost six figures and a dev team.

“I think probably the most powerful thing that we’ve launched is creating an MCP server into our own reservation software that allows AI tools to go and access tour information, departures, availability, pricing, and communicate all that in live time with customers through web chat.” Drew Fortner

What you’ll hear

How Alaska Shore Tours measures AI spend against minutes saved, and why the dashboard keeps landing around a 9,000% return.

Why 80% of their customers book within 24 hours of first finding them, and what that does to top-of-funnel spend.

The one ad message that beat everything else in their Google Ads headline testing.

How a mid-size reseller competes against the cruise ship for the same tour.

Listen or Watch The Episode

‍

SEO & GEO

Your ChatGPT Traffic Went to Direct

The Gist

→  ChatGPT dropped its UTM tags on October 9th, so GA4 logged a 22% to 57% collapse that never happened.

→  OpenAI called it a bug and shipped a fix on the 15th. Annotate October 8th in your reports before anyone reads it as lost traffic.

→  Most ChatGPT clicks still arrive without a referrer, so the tag is the only thing counting them.

ChatGPT stopped adding UTM tags to its links and stopped passing a referrer, so that traffic landed in Direct instead and GA4 showed it as a collapse. Graphite put the drop somewhere between 22% and 57% across the sites it tracks, while total traffic on those sites held steady.

Nobody lost the visitors, the label just came off them.

Adobe’s Cedric Huesler compared the week before against 9 to 13 October across roughly 500 pages where ChatGPT drove at least 5% of visits. ChatGPT visits fell by about 31,000 a day while Direct rose by about 26,000, and the new Direct visitors behaved like ChatGPT visitors, with first-visit rates climbing toward the ChatGPT norm.

The change arrived with GPT-6, and it hit inline links in the answer text while links in the Sources panel and shopping links kept their tags. Graphite found that 95% of mobile ChatGPT visits to its sites relied on the tag alone, which makes app traffic the most exposed.

OpenAI confirmed to Huesler on October 14th that it was a bug, and a fix began rolling out on the 15th. By that morning, tagged clicks were back to roughly 95% to 100% of the same hour a week earlier across the domains he tracks. Not everyone saw the fix immediately, and most ChatGPT clicks still arrive without a referrer, so the tag is what you count.

What to do about it. Annotate October 8th in every client report before anybody reads next month’s numbers as a loss. Then go page by page and compare Direct against ChatGPT, because pages that used to get ChatGPT traffic and suddenly gained Direct are your answer. Check your homepage too, since Peec AI found ChatGPT putting a direct URL into 42% of responses during the window, up from 8 to 10%, with 54% of those pointing at home pages.

There’s a version of this that keeps mattering after the bug is fixed. A channel that only exists in your reporting because a vendor chooses to tag it is a channel you don’t actually measure. Gemini added UTM parameters this month, ChatGPT removed them by accident and put them back, and neither of those was your decision.

Sources: Southern, Search Engine Journal  ·  SEJ on Gemini UTMs

‍

TRAVEL TECH

AI Agents are Going to Write the Commercials Rules - If You Don't First

The Gist

→ When Skift asked Meta’s Muse for a New York hotel, it opened a browser and picked between Expedia and Hotels.com on its own. The airlines in the same search had a signed integration with Duffel, which shows the benefit of integration into the AI agent.

→ Google’s agentic hotel booking keeps the supplier as merchant of record but keeps the conversation and all the details that produced the sale.

→ AI drives well under 1% of Booking Holdings room nights, which makes waiting to see how this develops seem sensible - but the commercial terms are being set now.

Every 2027 commerce budget has a line for AI, and most are sized against the bookings AI drives today, which is well under 1% of Booking Holdings room nights by the company’s own account. Bernstein’s Richard Clarke put AI referrals below 2% at both Booking and Marriott.

That makes waiting look prudent, but Skift’s argument is that volume is the wrong input, because the commercial positions harden while volume is still low.

The example that makes it concrete: when Meta’s Muse books a flight, it runs through Duffel’s direct airline connection, with loyalty numbers and cancellations supported. When Skift asked the same agent for a New York hotel room, it opened its own browser and chose between Expedia and Hotels.com unprompted.

The airlines had a defined commercial relationship with the agent. The hotels were shopped by software they never signed with.

The question isn’t whether agents matter yet, it’s whether you’re in the transaction by agreement or by default.

Integration isn’t free either. Google’s agentic hotel booking in AI Mode keeps the hotel or OTA as merchant of record, which sounds generous until you see what travels with it: name, payment, occupancy, property. The conversation that produced the booking stays with Google. Skift’s advice is to price that as distribution access with almost no customer data attached.

What we’d add. Two weeks ago we covered Meta and Sierra drafting the Personal Agent Protocol with Shopify, Stripe, Walmart and no travel company. This is the same story with a budget cycle attached. Standards and commercial terms are both being written this year, by whoever is in the room, and in travel that has so far meant airlines more than hotels.

The practical version for anyone sizing a 2027 line: the question isn’t how much AI booking volume to plan for. It’s whether your availability, your rates and your cancellation terms are in a state an agent can transact against, because the agents are already shopping you either way.

Sources: Skift

‍

PAID MEDIA

Google Is Handing Back Channel Controls

The Gist

→  New Channel Prioritization and Budget Allocation controls let you rank channels and cap spend on the ones you don’t want, including 0% inside Performance Max.

→  Every channel defaults to High priority, so doing nothing is itself a choice.

→  In February 2027 Google retires the existing Display allocation behavior. Search campaigns with Display enabled will shift spend unless you set Display to Low priority at 50%.

Google Ads is rolling out two controls over the coming months across Search, Shopping, Performance Max and Demand Gen. Channel Prioritization lets you mark channels as high or low priority, and Google funds the high ones first. Budget Allocation lets you cap what the low-priority ones can take, as a percentage of daily budget.

The setting that matters is zero, and Google confirmed 0% is valid, including inside Performance Max, which is the first real answer to anyone who has spent two years asking where their PMax budget actually went. You still can’t separate Search, YouTube, Discover, Gmail and Maps inside PMax, because Google groups its owned inventory together, but you can now switch off the groupings you don’t want.

Caps are best-effort rather than hard limits, which is worth knowing before you build a plan around them.

For travel this lands on a familiar argument. Search Partners and the Display Network are where a lot of us have quietly suspected budget goes to die, and the honest answer has always been that we couldn’t prove it or stop it without excluding the channel outright. A 20% cap is a test you can actually run.

The February problem. Right now, if you run Search with Display enabled, Google quietly favors Search inventory over Display. You didn’t set that up, it’s just how it works. In February 2027 Google retires that hidden preference, and because the new controls default everything to High priority, Search and Display get treated as equals unless you say otherwise. Google’s own advice for keeping things as they are is to set Display to Low priority with a 50% allocation, and there’ll be a one-click migration when the controls launch.

It’s unusual for a platform change to hand you more control and punish you for ignoring it. If you run Search with Display enabled, that’s a diary entry for January rather than February.

Sources: Osmundson, Search Engine Journal

‍