The short version
An AI assistant recommending a tour is making a bet. It cannot inspect your operation, call your office, or check whether Thursday's departure actually runs. It has to infer reliability from whatever evidence it can reach — and when the evidence is thin, it defaults to whoever is already famous.
A Trust Score is a way of making reliability legible to a machine. Tixxly's is a weighted composite of four operational factors — cancellation rate, response speed, booking success rate and reputation sentiment — normalised to a single score between 0 and 100. This post explains what each factor measures, how the weightings work, and what to do if yours is weak.
If you have not yet measured where you stand, start with how to audit what AI assistants say about your tours.
Why do AI assistants need a trust signal at all?
Because without one they fall back on fame, and fame is a poor proxy for quality in a market of small operators.
Researchers have named this the incumbent advantage: language models systematically recommend well-known brands more often than others, independent of merit (arXiv). For a category where the best cooking class in a city is often run by three people with no marketing budget, that is a structural problem. The model is not choosing the best option. It is choosing the most documented one.
The same research points to the way out. When tested against different persuasion techniques, models showed 50–73% breakthrough rates for credibility signals like authority and social proof, but only 10–13% for marketing language such as scarcity and anchoring (arXiv). Models largely ignore sales copy. They respond to evidence.
That is the whole argument for a trust metric. You cannot persuade an assistant to recommend you. You can give it verifiable reasons to.
How is a Trust Score actually calculated?
Four factors, each scored 0–100, then weighted into a single composite.
| Factor | Weight | What it measures | Why it carries that weight |
|---|---|---|---|
| Reputation sentiment | 30% | Volume, recency and tone of reviews across platforms | Reviews are the strongest single input to AI recommendations — 71% of AI travel responses use first-party reviews as a source (CMSWire) |
| Cancellation rate | 25% | Share of confirmed bookings the supplier later cancels | The clearest signal of operational reliability, and the one travellers feel most directly |
| Booking success rate | 25% | Share of confirmed bookings fulfilled without issue | End-to-end delivery is the actual product. Everything else is a prediction of this |
| Response speed | 20% | Median time from booking request to supplier response | Fast responses convert and indicate an actively managed operation |
The composite is straightforward: each factor is normalised to 0–100, multiplied by its weight, and summed. A supplier with a 40 reputation score and 95s across the operational factors lands very differently from one with the reverse — which is the point. The weightings encode what actually drives an AI recommendation, not what is easiest to measure.
Reputation sentiment (30%)
Not just star ratings. Volume, recency and tone, read across platforms and deduplicated.
72% of travellers always or frequently read reviews before deciding on things to do (TripAdvisor research), and recency carries real weight — one study found 85% of respondents are more influenced by recent reviews than older ones (IJSDR). Academic work on tourism reviews finds recency signals a business's current service quality rather than its historic reputation (National Library of Medicine).
Assistants apply the same logic, and they weight third-party corroboration far above anything you say about yourself. Owned brand content sits at the bottom of the citation hierarchy (Pepper). The gap between platforms is stark: Perplexity referenced review-based sources in 100% of travel responses tested, against 58% for ChatGPT and 56% for Gemini (CMSWire). Forty recent reviews across independent platforms will move a Trust Score more than any amount of homepage copy.
Cancellation rate (25%)
Supplier-initiated cancellations only — a traveller changing their plans is not your fault and is not counted.
This is the factor that most directly damages a recommendation. A cancelled booking is a traveller who acted on an AI suggestion and got let down, and it is increasingly visible to the platform that made the suggestion. As assistants move from suggesting to transacting, a supplier who cancels is a liability to the assistant's own credibility, not just to the traveller's holiday.
Booking success rate (25%)
The proportion of confirmed bookings that complete without an issue — no no-show, no refund, no failure in the handoff.
The most concrete factor, and the hardest to fake. If an assistant sends travellers to a supplier and the experience does not get delivered, that is direct evidence the recommendation was bad. Suppliers whose bookings complete reliably are simply safer to surface.
Response speed (20%)
Measured as the median time between a booking request and the supplier's response, not the mean — one holiday-week outlier should not define your score.
The average travel booking enquiry gets a reply in 42 hours (Gnosari). Most decisions are made in the first few. Leads contacted within five minutes convert at roughly 21 times the rate of those contacted after thirty (LeadAngel, via Gnosari).
That was already a revenue problem. In an AI-mediated market it becomes a visibility problem too. When an assistant checks live availability and gets a slow answer or none, it has learned something about you that it will act on.
What the Trust Score does not measure
Two things worth being explicit about.
Data quality is a prerequisite, not a factor. Structured, machine-readable product data — duration, meeting point, price, cancellation terms, live availability — is what makes you readable in the first place. Get it wrong and there is no Trust Score to weight, because the assistant cannot describe you accurately enough to recommend you at all. Research on how models select brands points repeatedly to the same mechanic: models name the businesses whose descriptions are most corroborated across independent sources, not the ones with the best copy (Lemniscate). Contradictory data is worse than sparse data. That work happens at the storefront layer, before the score exists.
Marketing does not register. Superlatives — "award-winning", "world's best", "unforgettable" — are exactly the register models discount. There is no keyword density to optimise when the output is a generated paragraph. There is no placement to buy. And a single glowing testimonial on your own site carries close to zero weight against sources a model can cross-check.
Five ways to improve your score
- Build review recency deliberately. The heaviest factor at 30%, and the slowest to move. A steady trickle of recent reviews across independent platforms beats a large but ageing pile on one. Ask at the end of the experience, not by email three weeks later.
- Cut supplier-side cancellations to near zero. Audit why they happen — minimum numbers not met, guide availability, overselling — and fix the operational cause. A cancellation is worth more damage than a slow reply.
- Instrument your booking failures. You cannot improve a success rate you do not measure. Find where attempts break, from checkout handoff to no-shows, and fix the top cause first.
- Cut response time to minutes, not days. Automate the acknowledgement even if a human handles the detail. The gap between five minutes and thirty is worth more than most marketing spend.
- Fix data contradictions before anything else. Pull up your own site and every OTA listing side by side and make duration, price, meeting point and cancellation policy identical everywhere. It is not scored directly, but nothing else counts until it is right.
Where Tixxly fits
Tixxly is AI infrastructure for travel commerce. We make suppliers legible to AI ecosystems — ChatGPT, Gemini, Perplexity, Claude and others — so their products can be found, understood and booked when travellers ask an assistant for recommendations.
Concretely, that means a published agent surface rather than a hope that crawlers work it out. Every active supplier gets an agents.txt policy file and an actions.json manifest describing exactly what an agent can do, plus endpoints for canonical product reads, live vendor-scoped availability lookups, and recording booking intent with idempotency handling. The technical detail is public in our agent documentation.
The Trust Score sits on top of that surface. It is available on the Pro tier, where suppliers see the composite score, the factor-level breakdown showing each contribution, and which factor is costing them the most.
We are not an OTA and we charge no commission, so suppliers keep 100% of booking revenue less card processing fees. Connect once, reach everywhere.
The assistants doing the recommending need a reason to choose one supplier over another. The suppliers who give them one will be the defaults.