Next-Gen Findability Discovery

Package holiday consumers will adopt AI for trip planning, but not if it costs them control or visual confidence. We got there by sitting with seven of them while they planned real trips.

loveholidays was evaluating how to evolve its search and discovery experience with AI. Before building anything, we needed to understand how consumers were already using AI tools in their planning: what they trusted, what frustrated them, and where AI got in the way.

Company	loveholidays
My role	Product Design Manager, research co-facilitator and observer
Researcher	Maria Narraway, Lead User Researcher
Method	Qualitative usability sessions: live platform walkthroughs, behavioural observation, attitudinal interviews
Timeframe	July – August 2025
Participants	7 package holiday consumers (1 pilot + 6 full sessions)

The brief: Evaluate how users discover and plan short city breaks with AI travel tools. Understand the attitudinal and behavioural gap between where AI search is today and what users need from it.

The platforms tested: loveholidays (structured search with keyword filters) and Mindtrip (freeform conversational AI for travel planning).

Three questions framed the study:

1. Attitudinal and behavioural patterns. How are users already combining traditional and AI-powered tools in their planning? What habits and mental models are they bringing into new interfaces?
2. Trust and booking confidence. How much do users rely on AI recommendations when making real booking decisions? What erodes or builds that trust?
3. Comparative performance. Where does loveholidays’ structured experience hold up against conversational AI, and where does it fall short?

Seven package holiday consumers. A mix of desktop and mobile users, household incomes ranging from £20k to £90k+, across employment statuses including full-time, homemaker, and self-employed. All were actively considering a holiday and had prior experience booking package deals.

5 of the 7 had already used AI tools, mostly ChatGPT or Gemini, for travel-related tasks. They weren’t coming to this cold.

Sessions ran approximately 60 minutes: a warm-up on current planning habits, a live platform walkthrough with both sites, structured preference questions, and a closing exercise imagining travel booking in five years.

I was present as an observer on every session. Maria Narraway led facilitation.

The volume of choice in package holiday search is the single biggest source of frustration. Participants described the research process as long, manual and draining. Several described a “give up” mindset, where they compromise on criteria they had called non-negotiable an hour earlier, to get out of the loop.

The problem is curation rather than filtering. Users don’t want to see everything available, they want the right things quickly.

What they wanted from AI was precision rather than breadth. “Find my perfect match” is a different product promise from “show me everything”, and the appetite for it is real. What’s missing is an experience that delivers it reliably.

Opportunity: Invest in curation-first AI design. Fewer, better-matched results, carrying signals about why they were surfaced, beat exhaustive inventories. Two-way conversation design, where the AI asks clarifying questions instead of returning one list, addresses this head on.

4 of 7 participants preferred loveholidays’ structured input over Mindtrip’s freeform AI chat when planning a package holiday. The reason was control rather than distrust of AI. Structured inputs showed how the results were reached. Freeform prompts felt like a black box.

Mindtrip was widely described as “the future”, impressive in concept and frustrating in use. Usability failures compounded quickly: itineraries appeared when users expected hotel lists, ATOL protection wasn’t surfaced, and there was no way to tell how results were generated or whether they were bookable.

A separate observation, from the six-participant analysed cohort: 5 of 6 didn’t spontaneously use loveholidays’ keyword filter during unprompted site interaction. When prompted, half weren’t confident the results were accurate. Feature placement and trust both need addressing.

Trust in the feature was a second, separate problem. One participant searched the keyword bar, a facility they reasonably expected most hotels to have, and the page returned nothing at all. They described the result as “worrying” and “concerning”, and moved from questioning the search to questioning the site.

Opportunity: Bridge traditional and AI interfaces rather than replacing one with the other. Put selection-based input alongside freeform prompting, and offer guided prompts and templates, because a blank text field is harder to face than a structured form. New AI patterns need deliberate onboarding.

In travel, visual content is how users make decisions. Participants described imagery, hotel photography in particular, as fundamental to even a superficial first assessment of a property. When AI tools returned text-heavy results, users disengaged.

Mindtrip’s text-first output was the most consistent source of frustration across the study. One participant described it directly: “I’d like something more visual — the image, the grid, the summary of the trip.”

Maps resonated strongly, particularly on desktop. Distance shown in context, “0.3km from the city centre”, was appreciated on sight and described as “really handy”. Spatial context does a job that lists can’t.

Opportunity: Imagery drives purchase intent, so it belongs inside the conversational interface rather than bolted on afterwards. Proximity information, map-integrated results and image-first grids are the gaps Mindtrip left open.

Users trust systems they can read. When filters are applied and results update in real time, confidence goes up. When AI generates a list with no visible logic, confidence collapses.

Two moments surfaced this clearly. The first: a participant working through Mindtrip’s ranked hotel list stopped and questioned whether the ordering was commercially neutral.

That is a fair question, and the interface gave them no way to answer it. The participant is describing how an unexplained ranking turns into a decision they didn’t consciously make.

A second participant noticed straight away that Mindtrip didn’t display ATOL protection for package holidays. One missing piece of familiar trust infrastructure made them hesitant to proceed. Missing signals cost as much confidence as unfamiliar ones.

Accuracy warnings (“results may not be accurate”) did the most damage of all. A disclaimer like that undoes the reason for using AI in the first place.

Opportunity: Get accuracy right before marketing AI features broadly. Show the reasoning where you can, and surface the trust signals users already look for (ATOL, Tripadvisor scores, pricing transparency) inside the AI interface.

Participants varied in how they approached AI prompting, and the difference tracked their mental model rather than their familiarity with technology.

Three patterns came out of the sessions. Some users wrote compressed, form-like prompts: “Barcelona, 3 nights, September, £200pp, 1km to city centre.” Others used natural, conversational language: “I want a 4-day holiday to Barcelona, staying by the beach, 4-star or above.” A third group used iterative follow-ups, asking things like “why are there no all-inclusives in Barcelona?”, and treated the AI as a domain expert rather than a search engine.

Two caveats sit underneath that diagram. The first is methodological, and one participant named it themselves: we watched people prompt while being watched, which is not how they prompt alone.

The second is an expectation gap. Participants assumed the system would infer intent they hadn’t spelled out, treating obvious context as understood rather than as something to specify.

The effort of typing and the effort of knowing what to type are both barriers. One participant went back to filter-based search because “ticking boxes is less effort”. Another, who has a cleft palate, found voice recognition failed to understand them, forcing repeated input and shutting them out of the feature.

Opportunity: The AI interface has to handle all three prompt styles as well as each other: compressed shorthand, conversational sentences, and iterative clarifying questions. Any voice recognition investment needs dialect and speech variation testing. Guided input templates lower the effort for users who don’t know where to start.

Four design constraints came out of the research:

1. Start with the familiar, enhance with AI. Users expect standard inputs (destination, dates, travellers) before advanced features. Replacing those inputs creates friction; building on them earns trust.
2. Mimic the human travel agent, digitally. Step-by-step, two-way conversations feel natural. Single-response AI that never asks a follow-up feels broken. Memory and contextual reasoning, as in “of the hotels you just showed me, which is closest to the beach?”, are expected rather than impressive.
3. Lead with visuals. Balance depth and simplicity. Users need enough to decide, not everything available. Images, maps and proximity signals carry the decision.
4. Curate, don’t overwhelm. Showing less with higher confidence wins. Tailored results that feel made for the user beat exhaustive inventories on both satisfaction and conversion intent.

On 28 October 2025 a new SRP card structure launched across all points of sale and both devices, built on the usability testing this study fed into and aimed squarely at the scannability problem Theme 1 and Theme 3 surfaced. Later readouts reported the experiment performing well.

That is the direct line from this research to shipped product. The study argued that users are drowning in undifferentiated results and decide visually, and the card is where both got addressed.

This is a transitional period. Participants who had built comfort with AI tools over months approached both platforms very differently from those encountering conversational travel AI for the first time. There is no single “user” to design for right now.

The safest route is several pathways: structured filters, guided prompts, freeform text and voice, with onboarding that helps users get the most out of whichever they pick. Hedge against the assumption that users are ready to abandon everything familiar.

The appetite is there. Several participants said, unprompted, that they would use AI over filters if it worked reliably. “If it can cut down the amount of time I spend clicking through pages, then I’d definitely use it.” Every one of those statements carried the same condition, and meeting it is the brief.

What’s still unresolved: This was 7 participants, so the findings are directional rather than statistically significant. A broader study, particularly one that maps where the wider loveholidays user base sits on the AI familiarity curve, would let us design for the distribution instead of the outliers. That is the next step.