The Summer Search Playbook

Part one of this series, “The Summer Discoverability Gap” (July 2026), measured how often ski areas appear when travelers ask AI about recreation in the mountains, revealing a visibility gap in the summer. Part two shows how AI models make recommendations, and how the ski industry can compete for the answer.

When a traveler asks ChatGPT where to go in the mountains for their summer vacation, the model makes a series of decisions. Understanding how that happens can help resorts position themselves to earn a mention.

AI models receive the prompt and respond in one of two ways, which, as an adjunct professor at the University of Utah, I like to compare to an open-book or closed-book exam.

In a closed-book exam, the student must answer based on what they know in that moment. An AI model can work in a similar way. 

Our research found that in roughly 20 percent of prompts about mountain recreation, AI responded by pulling from its own knowledge. Just like that student, the model learned by reading its textbook—the entire history of the internet.

In an open-book exam, the student can check their notes. Again, AI models can work the same way. In about 80 percent of the prompts we measured, the models opened the book and searched the internet for the answer. 

By studying that behavior, and exactly what AI models read, we can understand how resorts should position themselves.

AI Search Methodology 

In a closed-book search, a resort’s website is a small drop in the ocean of content on which a model has been trained, and that is part of the summer challenge. There is far more content about Mammoth Mountain, Calif., as a ski resort than as a summer recreation hub.

With over 300 million visitors per year to U.S. national parks, the models have trained on far more mentions of Yellowstone National Park than the Cowboy Coaster at Snow King Resort, Wyo. AI naturally assigns a higher probability to topics most prevalent in its training data. The more frequently a topic appears, the more likely it is to surface in a result.

That ocean of training content has to stop at some point, though. Most AI labs cut off training data anywhere from six months to a year in the past. The models you use today, like Claude Opus 5.5 or ChatGPT-6, don’t know what is on your summer event calendar this year, for example. They must search to find it.

The model decides for itself when it needs to search for more relevant information beyond its training data. The more an answer depends on current, local or specific facts, such as dates, events or a specific place, the more likely the model is to search. Centium measures that Claude searches the least when recommending travel destinations, typically below 50 percent of the time, while models like ChatGPT and Gemini search in more than 80 percent of prompts.

Query fan-out. When an AI model searches for more information, it performs a method of research called query fan-out. When you ask for the best hiking trail near Lake Tahoe, the model doesn’t search that phrase. It breaks the question into its own searches, runs them, and writes an answer from what it found. 

AI Models Seek Validation

AI has to rely on what others say online to draw a conclusion about an experience. And across every industry Centium measures, the most consistent behavior by AI models is validation-seeking. Before a model recommends anything, it looks for evidence that somebody credible has already judged it best. Roughly four in five prompts trigger at least one search containing the word “best” or “top.”

Mountain travel is no exception. 

In May 2026, SAM and Centium developed 345 brand-neutral prompts, which Centium ran across ChatGPT, Claude, Gemini, Perplexity, and Grok, for 1,725 AI responses across eight regional summer categories, seven cross-regional summer themes, and eight winter benchmarks. The prompts generated 7,652 citations across 2,668 domains and 1,677 fan-out searches from ChatGPT and Gemini. No prompt named a resort.

Nearly half of the fan-out searches in this study contained “best” or “top,” though some of that was a result of the prompt language, which often asked for “the top things to do” in a region in summer or the “best places” to go. Even so, what matters is what those searches found. Only 17 percent of the pages the models cited were ranked lists.

Share of “Best” or “Top” List Citations, By Model

Model	This Study	Centium Portfolio
Gemini	21 percent	41 percent
Perplexity	16 percent	42 percent
ChatGPT	6 percent	19 percent

The models found roughly half as much ranked content about mountain travel as they find in other industries tracked in Centium’s portfolio. They went looking for a ranking, mostly did not find one, and settled for whatever roundup they could locate. That is how a Marriott travel guide ended up the most-cited page in the study, and a travel insurance company’s blog post as the second.

The industry has an answer in winter. SKI magazine releases its Reader Resort Rankings, and OnTheSnow gives the models what they want in one place with rankings by state, snowfall numbers, lift and trail counts, and reviews. OnTheSnow is the most-cited ski-specific source in this study for winter prompts, but nobody has built the summer equivalent.

Abundant Guides 

While AI searched hardest for rankings, this was not the type of content it used most. Guide-style pages were 22.6 percent of everything cited, including things-to-do lists, attraction roundups, visitor guides, and itineraries.

The irony is that the models almost never use the word guide. They ask instead for things to do (which appears in 6 percent of searches), and for attractions, itineraries, and trips (each around 4 percent).

The guides AI cited most came from GetYourGuide, the trail platform Komoot, the National Park Service, and Tripadvisor, followed by many destination marketing organizations. The single most-cited guide in the entire study was a SKI magazine summer resort guide to the Northeast, referenced 19 times.

Resort-owned sites do get cited as guides, too. Whistler Blackcomb, B.C., Mammoth Mountain, Calif., and Tremblant, Quebec, all rank among the most-cited guide publishers on the strength of their own things-to-do pages. This is the one type of content where your own website genuinely competes.

Need for New Content

One of the clearest findings in Centium’s research is that AI chases recency. More than half of all prompts include at least one search naming a specific month, season, or year, with the models working “June” or “2026” into searches they write themselves. The standard advice that follows is to date your content and keep it current.

Remember that ocean of content the models already trained on? They know it, so when they search, they are seeking new information.

One-third of all URLs cited in the study contains a year, season or month. A blog post titled “Best Things to Do in Whistler This July,” for example, with your activities at the top, plays directly into how the models search.

Terms the Models Search For

In winter, 64 percent of fan-out searches contain a winter activity term, and that term is almost always ski. The activity is a given, and the model’s only real question is where.

In summer, the searches split across whatever the traveler asked about: hiking, trails, festivals, biking, spas, scenic lifts and gondolas, zip lines, coasters, dining, rafting, kayaking, and climbing. The models search in the traveler’s own words. When a prompt named no activity, the models added one to only 10 percent of their searches, most often hiking. In winter, they added a ski term to 29 percent of searches even when the prompt never mentioned skiing.

The models also stop looking you up by name. Searches naming a specific property fall from 38 percent in winter to 20 percent in summer.

Put those together and a single webpage titled “Summer Operations” with just ticket information has no chance. It matches few of the search terms that the AI models use. Instead, widening your footprint to have a page for each activity you offer on your website will increase your chances of surfacing. 

Pages on Your Website AI Read

We isolated every citation in the study that pointed at a ski area’s own website and classified each page.

Resort Website Citations, By Page Type

Page Type	Winter	Summer
Homepage	32 percent	14 percent
Activities and attractions	18 percent	21 percent
Events and calendar	5 percent	22 percent
Season landing page	1 percent	14 percent
Winter or ski product page	20 percent	1 percent
Mountain biking	0 percent	10 percent
Dining	0 percent	4 percent
Spa and wellness	0 percent	2 percent
All other pages	25 percent	12 percent

In winter, AI cites your front door. In summer, homepage citations fall from 32 percent of resort-site references to 14 percent. In winter, the homepage answers the question “where can I ski in Michigan?” In summer, the model needs a page about the thing it asked about, and if you do not have one it will cite somebody who does. 

Your events calendar is the most-cited page you own in summer, rising from 5 percent of resort-website citations in winter to 22 percent. It is the least glamorous page most ski areas maintain, and event terms appear in 7 percent of summer searches. Keep it current even after an event passes because when no newer page exists, the models continue to cite last year’s. In our study, for example, the models cited 2025 summer event pages from Big White and Sun Peaks, both in British Columbia, to answer 2026 festival questions.

Attach Yourself to a National Park

Part one of this report showed national parks taking seven of the ten most-named places in the study, and resort visibility dropping as resorts face more competition for outdoor activities in the summer.

Of the 429 summer answers that named a national park, 208 also named a ski area. Parks and ski areas share the answer roughly half the time. When a model names a park in its answer, the likelihood of a resort mention drops from about 60 percent to about 50 percent, but resorts are not shut out. What the model read matters more.

When a model builds its answer from the national park’s website, ski resort mentions fall dramatically. When it sources third-party regional content, resorts appear—in vacation rental roundups, a hotel brand’s travel guide, family travel blogs, and activity platforms like Komoot. We counted 219 pages in this study that pair a park with a town, a resort, or an activity, from “things to do in North Cascades National Park” to “best ski resorts for families” on a Rocky Mountain travel site.

You are not competing with Yellowstone. You are trying to appear in a regional article that mentions both, and your destination marketing organization is the shortest path there. Banff and Lake Louise, Alberta, publish content naming Banff National Park and the three ski areas within it together. In summer, five of the 13 answers that cited this DMO content named a Banff ski area, almost always for its summer gondola. 

What This Means for Ski Resorts

Given what we know, here are a few things ski areas can do right now to improve their AI discoverability, particularly when it comes to summer-focused searches:

• Publish a ranked list for your region, or pitch one to a regional outlet.
• Date your content by season, and keep it current.
• Build one page per activity, not a single summer page.
• Write guides, and use headlines like “things to do.”
• Get into the regional articles that name your nearby national park.
• Enhance your events calendar and ensure AI can read it.