The short version
Tour and activity suppliers pay OTAs 20–30% commission, and a large part of what that money buys is ranking — position in the OTA's own search results. When a traveller asks an AI assistant for a recommendation instead of browsing a marketplace, that ranking is not what decides the answer. The commission still comes out of the booking; the visibility it used to buy does not transfer.
There is also evidence that AI agents actively discount paid placement. In randomised experiments across Claude, GPT-4.1 and Gemini, a "sponsored" tag cut a product's selection probability from a 10% baseline to as low as 1.8%, while an unpaid platform endorsement raised it to as high as 72.7% (Allouah et al., arXiv).
Meanwhile the AI channel's own take rate is collapsing. OpenAI charged merchants 4% on in-chat checkout, and withdrew the feature roughly five months later after around a dozen merchants adopted it (Forbes).
The conclusion is not that commission is unfair. It is that commission is becoming a poor way to buy demand.
What does 20–30% commission actually buy?
Not just listing. Listing is close to free. The commission buys position.
Viator's Accelerate programme lets suppliers bid effective rates upward in exchange for ranking visibility. GetYourGuide tiers commission by category and supplier performance, with new suppliers commonly starting near 30% (Strathcode, Kong).
| Channel | Supplier take rate | What the rate buys |
|---|---|---|
| Viator | 20%, rising to 25–30% via Accelerate bidding (Strathcode) | Position in Viator search |
| GetYourGuide | 20–30%, tiered by performance (Strathcode) | Position in GetYourGuide search |
| Klook | 15–35%, typically 20–25% in APAC (Kong) | Position in Klook search |
| Amazon marketplace | 15–25% (Stellagent) | Position in Amazon search |
| ChatGPT Instant Checkout | 4% — withdrawn March 2026 (TechCrunch) | Transaction processing, explicitly not ranking |
| Google Universal Commerce Protocol | Free to implement (UCP Hub) | Nothing — it is a standard, not a marketplace |
The pattern in that table is the story. Every channel where the take rate is high is a channel that sells position. Every channel built for AI has a take rate at or near zero, because it is not selling position at all.
Why does paid placement stop working on AI assistants?
Two reasons, and they are different in strength. It is worth separating them.
The structural reason is certain. Paid ranking inside an OTA moves you up that OTA's results page. If a traveller never loads that page — because they asked ChatGPT or Gemini and acted on the answer — the ranking you paid for was never rendered. You are buying position in a shop window that fewer travellers walk past each year. This is arithmetic, not speculation.
The behavioural reason is strong but directional. Researchers at Columbia and Yale ran the first randomised, causal experiments on AI shopping agents, published as the ACES framework. Testing Claude Sonnet 4, GPT-4.1 and Gemini 2.5 Flash, they found agents "consistently penalize sponsored tags while rewarding platform endorsements" (Allouah et al.).
The size of the effect is what matters. In the API-style interface — the closest analogue to how an AI travel agent would read a product feed — a product with a 10% baseline selection probability moved as follows:
| Signal applied | Claude Sonnet 4 | GPT-4.1 | Gemini 2.5 Flash |
|---|---|---|---|
| Baseline | 10.0% | 10.0% | 10.0% |
| "Sponsored" tag | 5.4% | 1.8% | 8.9% |
| Unpaid "Overall Pick" endorsement | 58.4% | 55.6% | 72.7% |
Source: ACES, Allouah et al.
For GPT-4.1, labelling a product as sponsored cut its chance of being chosen by more than 80%. An unpaid endorsement raised it more than fivefold. The researchers confirmed the penalty persists in newer models — Claude Opus 4.5, GPT-5.1 and Gemini 3.0 Pro Preview all show negative sponsored coefficients — so this is not a quirk that model updates are erasing.
The honest caveat: that experiment tested an explicit "sponsored" label in a retail setting. OTA boost bidding is paid placement, but it is not labelled as such when a feed is exposed to a model. So treat this as directional evidence about how agents weigh commercial signals, not as proof about Viator specifically. The structural argument stands on its own regardless.
Do travellers want AI recommendations to be paid for?
No, and emphatically so. In a Harris Poll survey of 2,180 US adults conducted in February 2026, 75% said they would trust AI agents less if recommendations were swayed by brand payments. The same 75% said they would trust the brands less too (Quad).
That is a hard commercial constraint on the AI platforms, not a soft preference. An assistant's entire value rests on the user believing the answer is impartial. Monetising the ranking destroys the product. Which is why OpenAI stated plainly that its merchant fee did not "influence ChatGPT's product results", and why Google shipped its commerce protocol free.
Discovery in this channel is being decided by the quality of your product data, not by your budget. One analysis found 83% of products in ChatGPT's shopping carousel came from organic Google Shopping results (Dataïads, citing Peec AI via Search Engine Land).
Why does this hurt tours more than most categories?
Because supplier dependence on the commission channel is rising exactly as the channel's mechanism weakens.
Arival's Global Operator Landscape, 4th Edition — 5,664 qualified operator responses collected between August and November 2025 — found OTA share of tour and activity bookings rose to 37% in 2025 from 33% in 2024, while direct operator website bookings fell from 29% to 25% (Rise Strategic Consulting, citing PhocusWire's reporting of Arival).
Four points of share moved from direct to intermediated in a single year. Every one of those points is a booking where the supplier pays 20–30%, does not own the customer record, and holds no first-party data to feed an AI assistant with. Suppliers are increasing their reliance on paid position at the precise moment paid position stops being the deciding factor.
What should suppliers actually do?
- Work out your true effective take rate. Headline commission, plus boost or accelerate bidding, plus payout timing fees, plus payment processing. Most suppliers quote the headline and pay considerably more.
- Stop treating OTA rank as the goal. Rank buys visibility on one surface. Budget for it accordingly rather than as the whole distribution strategy.
- Make your own product data machine-readable. Structured schema on every product page, consistent pricing and availability, one canonical source of truth. This is the input AI assistants actually read.
- Fix contradictions across sources. If your site, your OTA listings and your Google profile disagree on price or duration, a model has no reason to trust any of them and will cite a competitor it can verify.
- Build the signals that cannot be bought. Review volume and recency, response speed, low cancellation rates. These are the closest real-world analogues to the endorsement signals agents reward.
- Own the checkout. The direction of travel across ChatGPT, Google and Perplexity is discovery in the assistant, transaction on the supplier's own infrastructure. Suppliers without a direct booking path are not addressable in that model.
- Measure AI visibility separately. It does not appear in Google Analytics as a channel you can optimise. If you are not testing what assistants say about you, you are guessing.
Where Tixxly fits
Tixxly is AI infrastructure for tour and activity suppliers, not an OTA. We charge zero commission — suppliers keep 100% of booking revenue, less card processing fees — because there is no ranking to sell. Suppliers pay a flat subscription: Free, Growth at $99/mo, Pro at $199/mo, or Infrastructure at $999/mo.
What the platform does instead is make a supplier's inventory legible to AI systems: a schema-marked storefront on an ai. subdomain, an open crawl policy for AI agents, structured product and availability data, and a Trust Score built from reputation sentiment, cancellation rate, booking success rate and response speed — the reliability signals that determine whether an assistant is willing to recommend you.
Connect once. Reach everywhere.