Inside the AI Ski Concierge: How We Built a Ski-Specific Assistant with Live Snow and Lift Data

    Inside the AI Ski Concierge: How We Built a Ski-Specific Assistant with Live Snow and Lift Data

    May 20, 2026
    By Son of a skibum

    # Inside the AI Ski Concierge: How We Built a Ski-Specific Assistant with Live Snow and Lift Data

    What does a ski traveler actually want from an AI assistant? Not a generic “best ski resorts” list. Not a vague answer about snow conditions “in the Alps.” They want to know whether the lifts are spinning at Zermatt, whether Val Thorens still has cold powder after last night’s storm, whether St. Anton am Arlberg is best for advanced skiers today, and whether it’s worth booking a family weekend in Whistler Blackcomb if the valley is warming up.

    That is exactly why we built the AI Ski Concierge: a ski-specific assistant designed to answer real skier questions using live snow, lift, weather, and resort knowledge rather than generic web guesses. It is part travel planner, part conditions checker, part resort expert. And unlike a general-purpose chatbot, it is built around the messy realities of ski season: changing snowpack, regional naming quirks, daily lift status updates, and the constant risk of stale data.

    If you’re an AI builder, product lead, or ski-industry watcher, this is a useful case study in how to design a vertical assistant that actually works. And if you’re a skier, this should give you a clearer picture of why ski planning is such a special use case—and why live data matters more than ever.

    Why ski search is harder than it looks

    On paper, “find me a good ski resort” sounds simple. In reality, ski queries are some of the hardest travel queries a model can handle well.

    Seasonality changes everything

    A resort can be incredible in January and a bad bet in April. A general model may know that the Alps are popular, but it cannot reliably infer whether a resort is currently open, how much snow is on the upper mountain, or whether valley runs are slushy. Skiing is one of the few travel categories where the answer changes daily and sometimes hourly.

    For example, a spring skier asking about Courchevel versus Chamonix does not just want “which is bigger.” They want to know snow reliability, altitude, aspect, lift access, and whether conditions suit piste cruising, off-piste, or family skiing. That is why we also point users to deeper editorial guides like the best French ski resorts and snow-sure resorts, but pair them with live resort data.

    Regional naming is messy

    Skiers rarely search by full formal names. They type “Zermatt,” “Arlberg,” “Whistler,” or “Trysil,” and they expect the assistant to understand context. They may also use local language names, shorthand, or regional clusters. A system must normalize aliases, map resorts to the correct records, and know when a query refers to a resort, a region, or a broader destination.

    That is especially important in Europe, where users often search by country and region. Someone asking about France may really mean the entire Alps, while a request for Italy might be narrowed to the Dolomites. In Scandinavia, a query about Sweden could lead to Åre, while Norway questions may focus on Trysil or Hemsedal.

    Prices change too often for static content

    Lift ticket prices, parking, lesson rates, and even booking windows can change fast. If an AI confidently invents a price, that is not a minor error—it destroys trust. Ski travelers are unusually sensitive to pricing because a mountain day can involve lift tickets, rentals, food, lodging, and transport. The wrong answer can alter the whole trip.

    That’s why we designed the AI Ski Concierge to be cautious about numbers, to prefer live sources where possible, and to avoid hallucinating specific pricing unless it can verify the data.

    The architecture: how the AI Ski Concierge works

    Our stack is intentionally simple in concept, but carefully layered in practice.

    1) Gemini for conversation and planning

    We use Gemini as the conversation layer because it is strong at natural dialogue, multi-step reasoning, and user intent parsing. Gemini helps the assistant handle open-ended questions like:

    • “Where should I ski this weekend if I want reliable snow and easy logistics?”
    • “Compare beginner-friendly resorts in Austria and France.”
    • “Is it worth going to Zermatt if I only have three days?”
    Gemini is also useful for summarizing resort characteristics into human-friendly recommendations. But we do not let the model freewheel on current conditions. That is where the data layer comes in.

    2) Perplexity Sonar for research fallback

    When a user asks a broader research question—or when we need supplemental context—we use Perplexity Sonar as a fallback research layer. This is especially useful for snow reports, industry context, resort news, and cross-checking details when a live API is incomplete.

    That research fallback is important because ski information lives across many sources and formats. A model might need to reconcile data points from a resort website, a forecast provider, and a destination guide. Sonar helps us keep the assistant grounded in recent information rather than relying only on training data.

    3) Regional live-data APIs

    Different regions require different data strategies, because ski infrastructure and data availability are not uniform.

    For North America, we use RapidAPI integrations to fetch live or near-live resort conditions, lift status, and relevant mountain metadata. In France, the ANMSM ecosystem is a valuable source for current mountain conditions and resort-side reporting. In Norway, fnugg provides a strong local source of live snow and resort information.

    This regional approach matters. A single global feed rarely covers the nuances skiers care about. The assistant needs to know where to look depending on the destination.

    4) A curated 675+ resort database

    The backbone of the system is a curated resort database with more than 675 ski destinations. This is not just a list of names. It includes structured fields for:

    • Resort aliases and alternate spellings
    • Country and region mapping
    • Elevation and terrain profile
    • Family, beginner, and expert suitability
    • Seasonality signals
    • Source references and update cadence
    This database lets the assistant identify whether a query refers to a resort page like Aspen or a broader region like the Rockies. It also helps route the request to the best data source and avoid confusing similarly named destinations.

    The hard problems: what we had to solve

    Hallucinated prices

    This was one of the first failure modes we noticed. General LLMs are eager to produce a number, even when none exists in the retrieved data. In skiing, that’s dangerous because lift ticket prices may depend on date, dynamic pricing, age bands, online booking, and even local promotions.

    Our solution: treat price data as untrusted unless it comes from a fresh verified source. If the system cannot confirm a number, it should say so. That sounds basic, but it is one of the most important trust safeguards in travel AI.

    Stale snow data

    Snow conditions are time-sensitive. A report that was accurate six hours ago may already be misleading after a warm wind event, overnight snowfall, or lift closure. That is why we use cache TTLs carefully and set different expiry windows for different data types.

    Weather and snow summaries may have a short TTL, while resort metadata can be cached longer. Lift status should be refreshed aggressively. A good ski assistant should know the difference between static resort facts and live operational conditions.

    If you want an example of why this matters, look at late-season destinations such as Val Thorens or springtime high-altitude skiing in Switzerland. Conditions can stay excellent, but the operational picture changes quickly.

    Name aliases and resort clustering

    Skiers use many names for the same place. “Arlberg” may mean the broader Austrian area, but the user might be specifically asking about St. Anton am Arlberg. Likewise, a question about “the 3 Valleys” should not be treated as a single resort page if the user is really asking about the region.

    Our resort database maps aliases, parent regions, and neighboring resorts so the assistant can understand intent rather than just keywords.

    Rate limiting and source resilience

    Live resort APIs are not infinite. Rate limits are real. So are outages and incomplete responses. We built the system to degrade gracefully: if one live source is unavailable, the assistant can fall back to another source, use cached data with a freshness warning, or answer with a narrower confidence range.

    This matters for user experience. A skier asking about “best snow tomorrow in France” should still get a useful answer even if one provider is slow. The assistant should never just fail silently.

    How ski-specific queries flow through the system

    Here are three examples of how the AI Ski Concierge responds.

    Example 1: “Is Zermatt a good choice this weekend?”

    1. Gemini identifies the query as a resort recommendation plus live conditions request. 2. The system resolves “Zermatt” to the resort record and pulls structured metadata from the database. 3. Live snow and lift data are checked through the relevant source path. 4. The assistant evaluates seasonality, altitude, and recent conditions. 5. It returns a concise answer with a confidence note and practical advice.

    A useful response might mention that Zermatt is typically a strong option for high-altitude snow reliability, especially when lower resorts are softening. It may also recommend checking the best Swiss ski resorts if the user wants alternatives.

    Example 2: “Compare family skiing in Åre and Trysil”

    1. The assistant detects a comparison intent. 2. It maps the request to Åre and Trysil. 3. It checks family suitability, beginner terrain, and likely operational consistency. 4. It generates a side-by-side summary and points to the broader family ski resorts theme.

    This is a good example of why ski assistants need domain knowledge. A family traveler does not just want snowfall statistics. They want lift layout, village convenience, and whether the resort is manageable with children.

    Example 3: “Where should I go for a snow-sure April trip?”

    1. The assistant recognizes a timing and snow-reliability query. 2. It prioritizes altitude, glacier access, and late-season track record. 3. It may compare Val Thorens, Zermatt, and high Alpine options in the Alps. 4. It cross-references our snow-sure resorts for April guide. 5. It answers with practical trip-planning advice, not just a list.

    What makes the assistant useful for real skiers

    The best ski AI does more than answer questions. It helps people make better decisions.

    Better resort matching

    A beginner needs very different advice from an expert chasing steep lines. Someone planning their first trip should be guided toward beginner-friendly resorts or the broader beginner ski resorts. An advanced skier may want off-piste destinations or the expert ski resorts.

    Better timing

    Should you go early season, midwinter, or spring? The assistant can steer users to our early-season ski resorts content or suggest the most snow-sure options when weather is marginal. That makes it far more useful than a generic travel chatbot.

    Better geographic comparisons

    Many users are still deciding between major ski regions. For them, comparison content matters. A question about the Alps versus North America can be linked to our Alps vs Rockies comparison, which helps frame terrain, lift style, and trip logistics.

    Why this matters for the future of travel AI

    The ski category is a perfect stress test for vertical AI.

    It combines:

    • Fast-changing real-world data
    • Regional fragmentation
    • Consumer trust issues around pricing
    • The need for structured and unstructured knowledge together
    • A user base that values precision
    That makes it a better AI product challenge than many “simpler” travel use cases. If you can make ski search reliable, you can learn a lot about how to build trustworthy AI for any weather-dependent or operations-dependent category.

    It also shows why generic LLMs are not enough. They can write a nice paragraph about skiing in France or Canada, but they cannot consistently answer whether the gondola is open, whether the base is melting, or whether the live snow report still matches reality.

    Practical advice for skiers using an AI assistant

    If you try an AI ski assistant, get the best results by being specific:

    • Include your travel dates.
    • Name the region or resort.
    • Say whether you care most about powder, beginner terrain, family convenience, or off-piste.
    • Ask for live conditions when timing matters.
    • Use follow-up questions to compare options rather than asking for a single “best” resort.
    For example, “I want a snow-sure four-day trip in March with easy airport access” will produce a much better answer than “Where should I ski?”

    Try the AI Ski Concierge

    If you want to see this in action, try the assistant at /ski-concierge-ai/. If you’re curious about the design choices behind it, we also explain the system in more detail at /how-it-works/.

    For skiers, the promise is simple: fewer stale recommendations, more live data, and better trip decisions. For builders, the lesson is just as clear: the best AI products are the ones that respect the domain.

    Actionable takeaways

    • Use live data for anything time-sensitive: lifts, snow depth, and weather.
    • Build alias mapping early so the assistant understands how skiers actually search.
    • Treat prices as high-risk data and verify before responding.
    • Use regional sources when global feeds are too generic.
    • Keep cache TTLs short for operational data and longer for static resort facts.
    • Add clear confidence and freshness cues so users know how much to trust the answer.
    • Pair AI responses with editorial content like our guides to create better ski decisions.
    The future of ski planning is not just smarter search. It’s ski-specific assistance that understands the mountain, the season, and the realities of live resort operations. That is the standard we built for the AI Ski Concierge—and the standard skiers should expect.


    Sources

    This article was researched using data from OnTheSnow, Skiinfo, Freeride, and other leading ski industry sources.