Centium Intelligence · Report No. 01
trends across 300,000+ query fan-out searches
When a user prompts an AI model, it does not search the exact phrase the user typed. It breaks the question down into its own searches, an approach called query fan-out. By studying how AI models search, Centium helps brands understand how to position themselves for what AI is looking for.
This report draws on 322,485 fan-out searches, measured across over 130 brands over a three-month span in 2026 and captured from ChatGPT and Gemini, the two models that expose their search behavior. On the average Centium dashboard, a single brand accounts for roughly 1,400 of these searches.
ai models are seeking validation
An AI model behaves like an evidence-based thinker. Before it recommends a brand, it looks for authoritative evidence that the brand is worth recommending: what experts and testers have already judged to be the best in a category, where that judgment is written down. Four in five of the prompts AI researches (80%) trigger at least one search for the best or top option. Zoom in on the individual searches and about 26% include the word “best” or “top,” the clearest signal of this validation-seeking.
Share of fan-out searches that include the word “best” or “top,” by sector
The rate climbs with how much a category is built around expert evaluation. Centium works with a large roster of outdoor and gear brands, and the outdoor industry turns out to be an ecosystem almost purpose-built for how AI thinks: expert gear reviewers, dedicated product testers, detailed specs, published rankings, and a steady stream of new releases to evaluate. It is no surprise that AI cites outdoor-industry media more heavily than any other segment we measure. OutdoorGearLab is the single most-cited source across our entire portfolio, and outdoor brands draw on specialist review media roughly three times as often as brands in other sectors.
Ranking is not the only way AI weighs options. In about 12% of prompts it runs a head-to-head comparison, an “X vs Y” search, and those concentrate in spec-driven categories like cycling, camping gear, and software, where buyers compare feature by feature. For an experience you cannot line up side by side, comparison barely registers.
A fair question is whether our own prompt phrasing drives this search behavior in the models. Because buyers ask for the “best trail shoes,” our prompts do too, and about half contain best or top. So we separated the prompts that use those words from the ones that do not. Prompts phrased neutrally still produced best-or-top searches 15% of the time, the model reformulating on its own. Prompts phrased the way a shopper would ask roughly doubled that, to 34%.
ai seeks reviewers, because it cannot test the product
An AI model cannot lace up the trail shoe or ride the bike, so when it evaluates a product it seeks out the people who have. More than a third of the prompts AI researches (38%) trigger at least one review search, and about one in seven of all its searches (14%) include the word “review.” Review-intent searches, where the model looks for tested, hands-on evaluations rather than marketing copy, are a distinctly physical-product behavior.
Ranking intent versus review intent, by sector
For gear, apparel, and consumer goods, roughly one in three searches seeks reviews. For destinations, hotels, and professional services, review-seeking nearly disappears, under 3% in travel and hospitality. Those categories lean on guides and rankings instead.
The sources reflect this. AI leans on expert product-testing publications whose entire model is buying the gear, testing it, and ranking it: OutdoorGearLab, GearJunkie, Switchback Travel, Treeline Review. In consumer goods the behavior is identical with different names, AI reaches for Healthline, Good Housekeeping, Allure, and Men’s Health, the health-and-beauty equivalents of the gear labs.
We also see AI models go to social media to find reviews. We have measured a citation to Reddit or YouTube in roughly 10% of Gemini’s citations and 17% of Perplexity’s, the two models that lean on them most, Reddit for candid user discussion and YouTube for hands-on video reviews the model reads through the transcript.
The scrutiny sharpens when we ask AI about a specific brand by name rather than a category. Review-seeking climbs to 21% of those searches, above the 14% baseline, and nearly a third of brand-name prompts send AI hunting for the downside: it runs searches like “is [brand] worth it,” “[brand] complaints,” “[brand] pros and cons,” even “[brand] scam.” AI vets a reputation the way a skeptical buyer would, looking for what is wrong before it decides what to recommend.
guides are a door into learning-heavy categories
Ranking and reviews are how AI evaluates a product. Guides are how it helps someone decide in the first place. About 15% of the prompts AI researches trigger at least one guide search, and they cluster in a telling place. When a category asks the buyer to learn before they choose, or to plan before they go, AI reaches for how-to content: guides, tips, and explainers.
Share of fan-out searches that seek a guide or how-to, by sector
Travel and tourism lead by a wide margin. A destination is not always a product you rank, it is a trip you plan, so AI hunts for “things to do” and “guide to visiting” content. Snow sports follow, where beginners have to learn how to choose gear before they can pick the best. Where the purchase is simple and familiar, guide-seeking falls away. For a brand in a learning-heavy category, the guide is a door AI walks through, and publishing genuinely useful how-to content is a way in that rankings and reviews do not offer.
ai is seeking recent content
AI models are already trained on a fixed snapshot of the past, so when they search the live web, they are looking for what is new. Almost half of all prompts (46%) include at least one search naming a specific year. Across the last three months that came to roughly 57,000 dated searches, about one in five of every search, and the year they name has been shifting toward the present.
Share of searches naming 2026, 2025 and 2024, by month
Through spring, searches naming 2025 and even 2024 outnumbered those naming 2026. By June, 2026 had climbed to the top. The pages AI cites are fresh to match. In our running-industry study, the median cited source was about four months old, and the sources AI leaned on hardest were fresher still.
For a brand, this means visibility is not a one-time win. The model refreshes constantly, so it decays unless you keep publishing current content.
the sources back up the search
We looked past the searches to the pages AI actually cited, to see whether the sources match the intent. They do, and the pattern runs deeper than a single format.
Share of cited URLs that are “best” or “top” lists, by model
Gemini44.6%
Perplexity35.7%
ChatGPT24.7%More than a third of every URL AI cites is a “best” or “top” list, led by Gemini at nearly half. Three things go further than the list. First, the pages are current: of the cited sources we could date, roughly six in ten were published in 2025 or 2026. AI is seeking fresh, recent content to broaden its knowledge base.
Second, AI does not always just stumble onto publications. Sometimes it searches for them by name. In our running study, one in every fourteen searches named a specific outlet, and Runner’s World alone was named in nearly one in twenty. Across the portfolio, Wirecutter and OutdoorGearLab are the outlets AI names most often. It will even reach for search operators, appending site:reddit.com to a query to pull candid discussion straight from the source. That is brand equity the model has internalized.
Third, how much AI leans on independent sources versus the brands themselves depends on what the category offers. In consumer products and other media-rich categories, the most-cited pages are overwhelmingly specialist media and expert reviewers. In B2B and industries with thin editorial coverage, AI has fewer third-party sources to lean on and goes more directly to the brands’ own sites. The rule of thumb: AI cites the media where the media exists, and falls back to the brand where it does not.
what this means for brands
- 01
Prioritize third-party validation and earned media.
AI leans more on independent publications, testers, and rankings than on your own website, which makes visibility as much a PR and earned-media challenge as an SEO one. Aim to be reviewed, tested, and ranked by the sources AI already trusts.
- 02
Activate your audience.
AI reads authentic reviews and community discussion, so encourage your satisfied customers to review you and talk about you where the model is looking, from Reddit and YouTube to the review sites your category relies on.
- 03
Keep the signal fresh.
Because AI chases recency, visibility decays without new material. Sustain a steady cadence of reviews, rankings, and coverage, yours and third-party alike, and date your content clearly so the current year works in your favor.
- 04
Act like a publisher.
Rethink your content strategy around what AI actually searches for: the guides, comparisons, and best-of content it hunts for, published on your own domain. The brands that earn their own site cited are running a genuine publishing operation, not posting product pages.
- 05
Measure what you cannot see.
All of this happens where your own analytics cannot reach. You do not know what you do not know: the searches AI runs about your brand, the sources it trusts, and what it surfaces when it looks you up. Centium uncovers exactly that, how AI searches about your brand and your category, so you can act on it rather than guess.
This report is built from the same fan-out data Centium tracks for individual brands.
Portfolio figures were captured from ChatGPT and Gemini, the two models that expose their search behavior, across more than 130 brands over a three-month span in 2026. Most prompts are unaided and category-level, never naming a brand; a subset asks about a specific brand directly, and those figures are labeled where they appear. A search is classified by the words it contains: “best” or “top” for ranking, “review” for reviews, “vs,” “versus,” or “compare” for comparison, and guide-style phrasing such as “guide,” “how to,” and “tips” for guides; recency is read from four-digit years, and the reputation grouping combines skepticism terms such as complaints and scam with worth-and-reputation terms.
We report two measures: share of searches, weighted by how often each search occurred, and prompt-penetration, the share of prompts whose fan-out included at least one matching search. To test whether our own phrasing drives the ranking signal, we compare fan-out from prompts that use ranking words against those that do not. Citation figures are drawn from current data across more than 225 brands and reflect the URL text and publish dates of cited sources.