A Comwrap Reply analysis of how agentic commerce is rewiring discovery and checkout for complex-journey operators, and the architecture decisions that decide who keeps the customer relationship.

Agentic commerce for complex-journey tour operators

Discovery is moving from Google to AI. For complex-journey searches, operators currently outrank OTAs in Google AI Mode, an unusual position that won't last, because the OTAs are investing heavily and will catch up. Underneath that, the commerce rails themselves are being rebuilt as an open standard stack (MCP, ACP, UCP, AP2) that, for now, keeps the operator as merchant of record. That second shift is the bigger one. It will run for years.

So there are two priorities. Get cited by AI now, while the window is open: published expertise on high-authority third-party publications (PhocusWire, Skift, Phocuswright, Arival, fvw|TravelTalk), under named human bylines, applying Princeton's GEO playbook of direct quotation, statistics, and primary-source citations. And be architecturally ready for whichever protocol and agent surface wins. The chatbot UI is commodity work. Any team can build one. The hard, durable asset is the data layer underneath: clean inventory APIs feeding a knowledge graph, with a RAG pipeline on top. Build that and you are protocol-agnostic and surface-agnostic.

A chatbot project is the wrong place to start. Unaided LLMs hallucinate inventory: invented destinations, fictional landmarks, wrong timetables. Air Canada is already legally liable for what its chatbot said. Without a knowledge graph grounding the AI in your actual products, any chatbot you build (or any third-party agent that connects to you) will misrepresent your business. The chatbot is replaceable in two years. The data architecture is not.

The handoff problem is the one most operators have not solved. When a customer spends 20 minutes with your AI assistant and then walks into your agency or calls your call centre, your systems must rebuild that conversational state in seconds. The "tour code" idea is decades old. The hard part is what sits behind it: clean inventory data, real-time availability, state portability across web/agency/call-centre platforms, identity continuity across channels. Most operators are not there yet.

Four moves, in order:

• Audit your current AI visibility (an afternoon's work, usually clarifying and unpleasant).
• Fund the GraphRAG pipeline. The knowledge graph keeps the AI honest. RAG makes it persuasive.
• Publish under named expert bylines on high-authority third-party publications, not your own domain.
• Get to Agentic Experience Readiness at the inventory layer: clean, normalized, fast APIs for the inventory you actually own. The least glamorous item, and the one that decides whether anything else works.

There is a quiet conversation happening in your customers' browsers that nobody on your team can see

A surgeon in Munich asks Claude to help her plan a three-week sabbatical through northern Vietnam: UNESCO sites, no internal flights, a guide who speaks French. A retired couple in Zurich asks ChatGPT for an expedition cruise to Svalbard that doesn't require a wetsuit but does include a glaciologist on board. A 38-year-old founder in Stockholm asks Perplexity for the three best operators running self-guided trekking in the Dolomites with luggage transfer and a vegetarian-friendly half-board.
In every one of these conversations, something is being decided about a tour operator's business. Whether the brand is mentioned. Whether it is recommended. Whether the operator the AI suggests is the curated specialist who actually owns the inventory, or a generic OTA that will commoditise the category and the margin.
This is not a forecast. It is happening this morning, in the highest-value segments of European travel, while operator e-commerce dashboards keep reporting what they have always reported.

The question for tour operator leadership is no longer whether AI is rewriting how complex trips get discovered and booked. It is whether you will be visible in that new layer, and whether, when an AI agent eventually closes a €15,000 expedition booking, you are the merchant who takes the order, owns the data, and keeps the customer.

At Comwrap Reply we work with European travel groups on the architecture, the publishing strategy, and the protocol integration this transition needs. The pattern is consistent. Two parallel shifts, with different time horizons. One is a finite window. The other is a structural rebuild that will run for years. Operators who handle both keep the customer relationship. Operators who handle neither lose it. Below is what we believe operators in the Explore & Experience and Pursue Passions segments need to know, along with the architectural choices that decide which side of the line they end up on.

The finite window: discovery is moving from Google to AI, and operators are currently winning

Search has changed shape faster than most travel groups have absorbed.
When Google places an AI Overview above the traditional links, the click-through rate on the top organic result drops by 58% [1]. Pew Research found that users click a traditional link only 8% of the time when an AI summary is present, versus 15% without one [2]. About 60% of US Google searches now end with zero clicks [3]. According to Phocuswright's latest tracker, 58% of active US travellers used generative AI for something in 2025, and 39% used it specifically for trip research [4]. Booking.com appears in 95.3% of all hospitality queries tested across ChatGPT, Gemini, Perplexity, and Mistral [5].

Here is the part most operator leadership has not yet been told: for curated tour and activity queries, operator-owned listings are currently outranking the OTAs inside Google AI Mode. Research by Arival and Propellic, published in September 2025, documented this clearly [6]. The discovery layer is being rebuilt in real time, and right now (for reasons of editorial depth, named expertise, and original content) operators with a real human point of view are unusually well-positioned.
This will not last. Booking.com, Expedia, and Tripadvisor went live as ChatGPT Apps in October 2025 [7]. Booking shipped its first agentic AI features the same month [8]. Expedia leadership has publicly stated that GenAI traffic converts to bookings at a higher rate than other channels. The OTAs know what is happening. They will close this gap. The question is whether your brand becomes the cited authority before they do, or after.

Generative Engine Optimization (GEO) is what fills this window. Princeton researchers tested nine content tactics across 10,000 queries [9]. Three move citation rates dramatically: adding direct quotation (+41%), adding statistics (+31–37%), and citing primary sources (+27–30%). Keyword stuffing was actively negative; it hurts AI visibility. The most counterintuitive finding for operators: lower-ranked sites benefit the most from GEO tactics, meaning category challengers can move faster than incumbents.

This is not a marketing tactic. It is a brand-distribution play with a hard expiry date. The operators who own the AI's default mental model for "small-group expedition cruises in Antarctica" or "self-guided cycling in the Loire" or "high-altitude trekking with cultural depth in Ladakh" will hold a distribution advantage for years. The ones who wait will be paying CPC to reach customers the AI is sending elsewhere for free.

The structural rebuild: the commerce rails are being laid for the next decade

Discovery is only the visible half of this transition. Underneath it, the commerce rails themselves are being rebuilt, at a pace the trade press has not yet conveyed.
Between September 2025 and January 2026, the major technology and payments companies launched four new protocols in rapid succession. Anthropic's Model Context Protocol (MCP) is now adopted by OpenAI, Google, and Microsoft, letting any AI agent talk to any business system. Stripe and OpenAI launched the Agentic Commerce Protocol (ACP) in September 2025, with Etsy, Shopify, Walmart, and PayPal as launch partners [10]. Google announced the Agent Payments Protocol (AP2) the same week, with sixty partners including Mastercard, American Express, PayPal, and Coinbase [11]. In January 2026, Sundar Pichai used the NRF stage to announce the Universal Commerce Protocol (UCP), endorsed on day one by Etsy, Wayfair, Target, Walmart, Adyen, American Express, Mastercard, Visa, PayPal, Stripe, and Zalando [12]. Visa reported "hundreds" of completed agent-initiated transactions by year end.
You do not need to follow the alphabet soup. You need to follow the direction.

These protocols share one important architectural choice: the merchant remains the merchant of record. The agent helps the customer choose. The merchant takes the order, the payment, the data, and the liability. That is what is on offer today, while the standards are being shaped. Whether the terms remain this favorable once the dominant agent platforms reach real scale is an open question, and history is unkind to merchants who arrive at standards bodies after the standards have been written.

None of this is settled, and none of it is risk-free. OpenAI quietly deprioritised its in-chat shopping checkout in March 2026, reorganising around its broader Apps strategy [23]. Senator Elizabeth Warren wrote to Sundar Pichai in February 2026 raising antitrust concerns about UCP. Citation visibility in AI engines is volatile: roughly 70% of brands lose visibility within six months without continuous reinforcement [24]. Cloudflare data shows AI crawlers extract content at ratios as high as 286,000 page-fetches per single referral click [25]. None of it is comfortable, and none of it is a reason to wait. It is a reason to build at the protocol layer rather than chase whichever consumer surface is fashionable this quarter, because the protocols persist even as the consumer surfaces churn.

The work to be architecturally ready for these protocols is no longer experimental. It is reference-architecture work. The operators who are ready to plug into whichever protocol stabilizes will spend the next two years composing new capabilities on top of a working substrate. The ones who are not will spend the next two years on integration projects while their competitors are building products.

The chatbot is the easy part. The data underneath is not.

Integrating with the rails is only half the work. The other half is what runs on top of them: what the AI actually knows about your business, and how it answers when a customer asks. That is where the next mistake usually happens.
Most operators hear "AI" and immediately ask their team to build a chatbot. This is the easy half of the problem, and the wrong place to start.
Building a chatbot UI in 2026 is straightforward. Any half-competent team can stand one up in a few weeks. Open-source frameworks exist. SaaS vendors sell pre-built widgets. You can vibe-code your way out of it. The presentation layer is a solved problem. Feeding that chatbot with clean, accurate, queryable data about your actual business is not.

Unaided large language models are catastrophic at the kind of inventory complex-journey operators sell. In September 2025, the BBC documented ChatGPT inviting tourists to the "Sacred Canyon of Humantay" in Peru, a place that does not exist. Travellers paid for transport to reach a road in the middle of nowhere. Layla, an AI travel app, told users about the Eiffel Tower in Beijing. A hallucinated cable-car timetable stranded a hiker on Mount Misen in Japan [13]. On the academic benchmarks, GPT-4 solved 0.6% of multi-constraint travel-planning tasks [14]. Independent testing found that 90% of AI-generated itineraries contained at least one factual error.

The legal exposure is no longer theoretical. In February 2024, the Civil Resolution Tribunal of British Columbia ruled that Air Canada was legally responsible for a chatbot that hallucinated a bereavement-fare policy [15]. It is now the leading global precedent on chatbot liability. Gartner forecasts more than 2,000 AI-related liability claims by the end of 2026 [16].

For an operator selling €5,000 cultural tours and €25,000 expedition cruises, the implication is simple. The chatbot is a commodity. The data the chatbot pulls from is not. And the data layer is where most operators are not ready.

What "ready" looks like, in concrete terms: clean, normalized, machine-queryable inventory feeding a knowledge graph (a relational map of destinations, accommodations, activities, transport segments, themes, certifications, seasonality, accessibility constraints, and combinability rules), with a retrieval-augmented generation pipeline on top that pulls in destination context, expert commentary, trip-designer notes, customer testimonials, and the editorial richness that turns a constraint set into a story worth €15,000. The knowledge graph keeps the AI honest. The RAG layer makes it persuasive.
The economics of this stack are favorable. Microsoft Research's GraphRAG architecture cuts AI compute cost per query by 26 to 97 percent versus unaided approaches [17]. LinkedIn deployed a knowledge-graph-backed AI to its customer-service workflow and reduced median issue-resolution time by 28.6% [18]. Amazon, Airbnb, eBay, and Booking.com have all published material on their internal knowledge graphs as the foundation of their AI-era architecture. This is not a research curiosity. It is the operating layer of every serious AI commerce business.

The strategic point matters more than the technical one. Build this architecture, and you are no longer dependent on which chatbot wins, which protocol stabilizes, or which agent platform reaches scale first. Whatever interface eventually talks to your customer (your own assistant, ChatGPT Apps, Gemini AI Mode, Comet, an agent surface nobody has launched yet) can pull from your clean data layer through whichever protocol turns out to be the standard. The chatbot you build today may be replaced in two years. The data architecture you build today is still useful in ten.

That is the asymmetry. The chatbot is the easy, replaceable presentation layer. The data architecture underneath is the hard, durable asset that pays back for years. Most operators are about to invest in the wrong half of that pair.

But architecture alone does not close a €20,000 expedition booking. The harder design question is how the AI assistant fits into the human-led sales motion that defines a complex-journey operator in the first place.

The handoff problem: from AI conversation to agency desk in seconds

The objection we hear most often from operator leadership is the right one. Complex journeys need a human touch. A €20,000 expedition is not bought by an autonomous agent on a Tuesday afternoon.
That objection is correct. It is also a false binary.
The customer evidence is clear. Skift's most recent research found only 2% of leisure travellers willing to let AI book on their behalf, against 80% of executives planning to deploy at scale [19]. Bain found that 50% of consumers are uncomfortable with fully autonomous purchases [20]. Klarna's CEO publicly reversed his AI-first customer-service strategy in May 2025 and rehired humans, calling quality human support "the way of the future" [21]. For complex, emotionally loaded, high-trust purchases (which is exactly what our segment sells), customers want both the AI's speed and the human's judgement, in whichever sequence they prefer.
The right architecture pairs an AI assistant that does what AI does well (exploration, configuration, constraint-satisfaction, availability lookup) with the human channels operators already own. Agency networks, call centres, trip designers. They close a high-trust decision and own the customer relationship.
The question is not whether to combine the two. It is whether your infrastructure can actually do it.

Consider what the customer expects. They spend twenty minutes on your website talking to your AI assistant, refining a Vietnam itinerary. They pause. They walk into your travel agency in Frankfurt the next afternoon. The agent should be able to ask for a reference, pull up the customer's full configuration in seconds, see what was discussed, what was ruled out, what is still open, and continue the conversation as if they had been in the room the whole time.

Most operators cannot do this today.
The idea of a booking reference or configuration code that travels with the customer is not new. Tour operators have used short references for decades. What has changed is what the reference now needs to carry. The customer arrives at your agency desk not with a postcode and a destination preference, but with twenty minutes of conversational state: preferences expressed, constraints established, options considered, options ruled out, open questions, party composition, payment readiness. Re-loading that state instantly across your web stack, your contracted agency desktops, your call-centre platform, and your back-office booking engine is an infrastructure problem most operators have not solved.
The underlying engineering is harder than it looks. The data layer has to support it. Customer state has to be persisted in a form any downstream system can re-load, not as a chat transcript a human has to parse. A travel agent in your contracted Frankfurt branch needs the same view of the customer's configuration as your call centre in Hamburg. The system of record has to be clear. Identity has to be continuous across channels. APIs have to be fast enough that the agent does not lose the customer to a coffee while the system loads.

A short, customer-facing reference (whether you call it a tour code, a booking ID, or a six-character handle) is the easy, visible part. The hard, invisible part is everything that has to be true behind it for that reference to actually mean something. Clean inventory data. Real-time availability. State portability between the AI conversation, the web checkout, the agency desktop, and the call-centre platform. Identity stitching across whichever channel the customer started in and wherever they want to finish. A single source of truth for who owns the booking at any given moment, and what permissions each channel has to modify it.

This is the kind of infrastructure work that does not produce a quarterly demo. It does, however, decide whether your operator can credibly tell a customer "start the conversation with our AI, finish wherever you like" or whether you stay locked into channel-specific experiences that lose customers at every handoff.

This is also the answer to a question we are increasingly asked: what does Reply actually build for tour operators in this space? The honest answer is the connective infrastructure. The clean inventory layer that lets your channels share state. The knowledge graph that prevents hallucination across surfaces. The API integrations that let an agency desktop re-load an AI conversation in seconds. The protocol-layer readiness that keeps your options open as the agent ecosystem matures.

What we recommend

The architecture is clear. The business case is durable. Some of this is time-bound, some is structural, all of it rewards early action. What remains is execution: four moves, in order.
First, commission an honest audit of your AI visibility. What does ChatGPT say when asked to recommend a small-group expedition operator in your category? What does Gemini surface for "best self-guided trekking, northern Vietnam"? What does Perplexity cite when asked to compare you to your three closest competitors? Most operator leadership in our segment has never done this. It takes an afternoon. The result is usually clarifying and unpleasant. It is the only basis on which a serious AI strategy can be built. Adobe's LLMO will do the job perfectly.
Second, publish like a media company under named human expert bylines, but publish where it counts. Your destination specialists, expedition leaders, cultural guides, and trip designers are the most defensible content asset you own. They are exactly the kind of authoritative, original-research voice that AI engines cite preferentially over corporate marketing copy. But the critical detail is where this content lives. Articles on your own domain carry far less GEO weight than the same expertise published on high-authority third-party publications. For European complex-journey operators, the publications that move the needle are PhocusWire, Skift, Phocuswright, Arival, and fvw|TravelTalk. AI engines weight third-party authority above self-published content by a significant margin. An expedition leader quoted in PhocusWire on the future of polar travel will be cited by ChatGPT in conversations about your category for years. The same expedition leader's blog post on your own site will not.

Third, fund the GraphRAG pipeline. Not as an "AI project". As a foundational data architecture investment with two layers that build on each other. The first layer is the knowledge graph: every product, component, constraint, and combinability rule, machine-readable and queryable. This is the layer that prevents the AI from hallucinating your inventory, your dates, your prices, or your availability. The second layer is the retrieval-augmented generation pipeline that sits on top, pulling in destination context, expert commentary, trip-designer notes, customer testimonials, photo metadata, seasonal observations, and the editorial richness that turns a dry constraint set into a trip a customer actually wants to buy. The knowledge graph keeps the AI honest. The RAG layer makes it persuasive. Microsoft Research's GraphRAG architecture combines both and cuts compute cost per query by 26 to 97 percent versus unaided approaches [17]. LinkedIn's production deployment of the same pattern reduced median resolution time by 28.6% [18]. This is the single piece of work that pays back across every other AI initiative for the next five years. It needs patient capital, because a GraphRAG pipeline does not produce a quarterly demo.

Fourth, get to Agentic Experience Readiness at the inventory layer. None of the first three moves works without it. Your tours, packages, and curated itineraries (especially the inventory you own rather than resell from third parties) must be clean, normalized, and accessible through fast, well-documented APIs. Clean inventory APIs are the substrate that everything else sits on. The GraphRAG pipeline pulls from them. The MCP servers wrap them. The agentic checkout protocols (ACP, UCP, AP2) transact against them. The cross-channel handoff between web, agency, and call centre runs through them. Every protocol you adopt on top of messy inventory amplifies the mess. Every protocol you adopt on top of clean inventory amplifies the asset. This is the most boring item on the list and the one that decides whether anything else pays back.

Note what is missing from this list of four: "build a chatbot." That omission is intentional.

The choice

Rafat Ali, in a Skift essay in late 2025, framed the choice for the industry more sharply than any analyst this cycle [26]. Do you want to be the destination travellers visit, or the invisible infrastructure that banks, airlines, retailers, and AI agents call when travel needs to happen for someone else's customers? For curated, complex-journey operators in the Explore & Experience and Pursue Passions segments, the answer is not really in dispute. The brand is the curation, the relationship, and the trust earned over a 15-day journey through territory the customer could not have navigated alone. None of that is at risk from AI. It is at risk from inaction.

The discovery layer is moving. The checkout layer is being standardised. Some of these changes have a finite window. Some are structural and will run for years. All of them reward operators who are ready. Operators who use this moment to get to Agentic Experience Readiness at the inventory layer, build a GraphRAG pipeline, publish for AI citation through third-party voice, and integrate at the protocol layer will have a real lead by 2030. Those who treat it as a 2027 problem will discover that 2027 is not when the change happens. It is when the consequences of having missed it become unavoidable.
None of this is speculative. The technology is mature, the architecture is documented, the protocols are open, and the empirical evidence (LinkedIn, Amazon, Airbnb, Microsoft, Princeton, Phocuswright, Bain, Skift) is substantial enough that no leadership team can credibly claim this is too early to call.
The harder question is the operational one: which of the four moves above lands on your roadmap before the end of Q3, and which gets quietly deferred to "next year".

Comwrap Reply is part of Reply Group, working with travel and commerce clients. Our practice focuses on Agentic Experience Readiness at the inventory layer, GraphRAG architecture for travel, and Generative Engine Optimization. If your team is wrestling with where to start, we are happy to share the reference architecture we use with our clients.

References

[1] Law, R. (December 2025). Update: AI Overviews Reduce Clicks by 58%. Ahrefs. https://ahrefs.com/blog/ai-overviews-reduce-clicks-update/

[2] Pew Research Center (July 2025). Google users are less likely to click on links when an AI summary appears in the results.

[3] SparkToro / Datos (2024–2025). Zero-click search analysis: 58.5% of US Google searches end without a click.

[4] Phocuswright (December 2025). Travel Forward: Data, Insights and Trends for 2026. https://www.phocuswright.com/Travel-Research/Research-Updates/2026/Travel-Forward-Data-Insights-and-Trends-for-2026

[5] Nokumo Research (2026). How Does AI Recommend Hotels? 450 Queries, 4 Models, 5 Countries. https://www.nokumo.net/en/blog/how-does-ai-recommend-hotels-we-tested-450-queries-across-4-models-to-find-out

[6] Quinby, D. & Visser, J. (September 2025). In AI Mode, Operator Listings Rule (for Now). Arival × Propellic. https://arival.travel/article/in-ai-mode-operator-listings-beat-otas/

[7] OpenAI (October 6, 2025). Introducing apps in ChatGPT: Expedia, Booking.com, Tripadvisor and more go live.

[8] Booking.com (October 9, 2025). Booking.com Debuts Agentic AI Innovations. https://news.booking.com/bookingcom-debuts-agentic-ai-innovations-adding-to-its-robust-suite-of-genai-tools-for-customers/

[9] Aggarwal, P., Murahari, V., et al. (2024). GEO: Generative Engine Optimization. KDD 2024, arXiv:2311.09735. https://arxiv.org/abs/2311.09735

[10] Stripe & OpenAI (September 29, 2025). Stripe powers Instant Checkout in ChatGPT and releases Agentic Commerce Protocol. https://stripe.com/newsroom/news/stripe-openai-instant-checkout

[11] Google Cloud (September 16, 2025). Announcing Agent Payments Protocol (AP2). https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol

[12] Pichai, S. (January 11, 2026). NRF 2026 remarks on the Universal Commerce Protocol. Google. https://blog.google/company-news/inside-google/message-ceo/nrf-2026-remarks/

[13] BBC Travel (September 26, 2025). When AI travel planning goes wrong: invented destinations, fictional landmarks, hallucinated timetables.

[14] Xie, J., et al. (2024). TravelPlanner: A Benchmark for Real-World Planning with Language Agents. arXiv:2402.01622. https://arxiv.org/abs/2402.01622

[15] Moffatt v. Air Canada, 2024 BCCRT 149 (Civil Resolution Tribunal of British Columbia, February 14, 2024).

[16] Gartner (October 2025). Top Predictions for IT Organizations and Users in 2026 and Beyond. https://www.gartner.com/en/newsroom/press-releases/2025-10-21-gartner-unveils-top-predictions-for-it-organizations-and-users-in-2026-and-beyond

[17] Edge, D., et al. (2024). From Local to Global: A Graph RAG Approach to Query-Focused Summarization. Microsoft Research, arXiv:2404.16130. https://arxiv.org/abs/2404.16130

[18] Xu, Z., et al. (2024). Retrieval-Augmented Generation with Knowledge Graphs for Customer Service Question Answering. SIGIR 2024, arXiv:2404.17723. https://arxiv.org/abs/2404.17723

[19] Skift Research (March 2026). Travel Brands Are Building AI Agents for a Consumer That Doesn't Exist. https://skift.com/2026/03/03/travel-brands-are-building-ai-agents-for-a-consumer-that-doesnt-exist/

[20] Bain & Company (December 2025). Agentic AI in Retail: How Autonomous Shopping Is Redefining the Customer Journey. https://www.bain.com/insights/agentic-ai-in-retail-how-autonomous-shopping-redefining-customer-journey/

[21] Klarna (May 2025). CEO Sebastian Siemiatkowski public statement on rehiring customer-service staff after AI-first reversal.

[23] CNBC (March 20, 2026). OpenAI's first crack at online shopping stumbled. It's preparing for the next wave. https://www.cnbc.com/2026/03/20/open-ai-agentic-shopping-etsy-shopify-walmart-amazon.html

[24] BrightEdge / AirOps (2025). Citation drift analysis: 70% of brands lose AI search visibility within six months.

[25] Tomé, J. (August 29, 2025). The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals. Cloudflare. https://blog.cloudflare.com/crawlers-click-ai-bots-training/

[26] Ali, R. (December 17, 2025). Famous vs. Indispensable: The Choice AI Is Forcing on OTAs. Skift. https://skift.com/2025/12/17/famous-vs-indispensable-the-choice-ai-is-forcing-on-the-otas/

primary-language-url