What AI Searches for Before It Answers
Your Query Fan-Out tab now classifies what every AI search was trying to find out, and shows which competitors turn up while the model is still searching.
the searches, sorted by purpose.
Before ChatGPT or Gemini answers a question about your category, it runs its own searches. Centium has captured those micro-searches for a while, and the Query Fan-Out tab listed them.
Two new cards sit above that list. Search Intent classifies every search into what it was trying to find out: a ranking, a review, a comparison, something recent, a guide, a purchase, alternatives to you, or a specific source. Competitors in the Searches crosses those searches against your competitor set, so you can see who the model pulled up while it was still deciding.
Search Intent
Seeks a ranked winner, the best or top option.
Search terms best, top, top rated, leading, highest rated
Sample data from the demo dashboard, with Ranking expanded. Prompts is the share of prompts that triggered at least one search of that intent, so the column does not total 100: a single prompt fans out into several searches with different purposes.
The classification is deterministic word matching, not a second model judging the first. That matters more than it sounds: the same search always lands the same way, so the mix is comparable run over run. Open any row and it shows the trigger words that fired, down to the year for a recency search or the domain for a site: one.
we measured it before we shipped it.
This came out of Report No. 01, where we measured 322,485 fan-out searches across more than 130 brands to see how the models actually research a category before recommending anyone.
The finding that started it: the models are not searching for you. They are searching for a ranking, a review, a comparison, a recent article. Four in five prompts trigger at least one search for the best or top option in the category. Whether your brand turns up depends on whether the pages answering those searches exist and mention you.
That was a portfolio-wide read. It said what AI does in general, and left every brand asking the obvious next question, which is what it does in my category. These two cards are that answer, computed on your own searches.
what the mix is telling you.
- A heavy ranking mix means listicles decide your category. The pages that rank brands are doing the work. Being absent from them is the whole problem, and amplifying the coverage you have is the fastest route in.
- A heavy review mix means the models want evidence. They are looking for someone who used the product. Owned copy will not satisfy that search; third-party coverage will.
- Source-seeking searches name the site AI trusts. When a search carries a `site:` operator, the model is going somewhere specific. That domain is worth being on.
- A competitor searched but never named is a near miss. The model considered them and left them out. When that is you, the coverage exists but is not convincing yet.
Both cards split category searches from brand-typed ones, so you can tell the difference between how AI researches your category and how it researches you by name.
questions, answered.
Eleven themes: ranking, review, comparison, recency, guide, purchase, alternatives, discovery, source-seeking, scrutiny, and branded. Every fan-out search is classified against all of them, so one search can carry more than one intent when it is doing more than one job.
By deterministic word matching, not by another AI model. The same search always lands on the same intent, which is what makes the mix comparable from one run to the next. Open any row and it shows the exact trigger words that fired, so you can check the call yourself.
ChatGPT and Gemini, the two that expose the searches they run before answering. Claude, Perplexity, Grok, and Google AI Mode do not surface their internal searches, so they contribute to your visibility numbers but not to the fan-out tab.
Because these are the brands AI reached for while researching, not the ones it named in the finished answer. A brand can be searched and then left out. That gap is worth reading: the model considered them.
No. Both cards are computed from the fan-out searches your dashboard already stores, so they appeared on existing data rather than waiting for your next update.
More guides
Which Writers AI Cites in Your Category
Two new features are live. Influential Voices names the people behind the pages AI cites, and Content Freshness shows how old that content is.
Expanded Sentiment Tracking in Brand Perception
Brand Perception shows the themes AI models use when they talk about you. Each theme now expands to explain what the models actually said, in their words, with sentiment attached.
Content Amplification: Own Your Wins
Media companies are blocking AI crawlers, and the coverage you worked hardest to earn can go invisible overnight. Content amplification keeps those wins where the models can still find them.