What Are Fanouts?
When an AI model answers a question, it may run several narrower web searches before composing its response. These behind-the-scenes searches are called fanouts. For example, a broad prompt about choosing software for a retail business might fan out into searches about pricing, integrations, reviews, alternatives, and retail-specific use cases. The wording of those searches reveals how the model interprets the original question and what information it decides to look for. The Fanouts page helps you answer questions such as:- Which searches do AI models run most often for my market?
- Which tracked prompts lead to the same search?
- Does a model introduce my brand or a competitor even when the prompt does not mention either one?
- Which words and concepts does the model add or remove when it rewrites a prompt as a search?
A fanout is a search query, not a citation or an answer mention. It shows what the model looked for. Use Prompts & Citations to see the sources it cited and the answer it ultimately produced.
Filter the Data
Fanouts opens to the last 30 days. Use the filters at the top of the page to narrow the analysis by:- Category — focus on one category or analyze behavior across all categories
- Topic — inspect the prompts grouped under a specific topic when a category is selected
- Tags — analyze a custom set of tracked prompts
- Time range — review a recent window or choose custom dates
- AI model — compare all available models or isolate one model’s search behavior
Read the Overview
Four summary metrics show the scale of the selected data:- Fanout searches — the total number of searches issued
- Answers with fanouts — the number of model answers that issued at least one search
- Searches per answer — the average number of searches issued by each of those answers
- Prompts covered — the number of distinct tracked prompts that produced fanout data
Explore Top Fan-out Searches
Top fan-out searches ranks the exact search strings AI models issued. Each row shows the models that used the search, the number of distinct prompts that led to it, and its total search count. Use this table to:- Sort by Search count to find the most frequently issued searches
- Sort by Prompts to find searches that many different questions converge on
- Turn on Branded only to show searches that name your brand or a tracked competitor
- Enter text in Filter searches to investigate a product, capability, problem, or competitor
- Export the current ranking for further analysis or reporting
- How many answers and distinct prompts issued the search
- Which AI models used it
- Other searches most often issued by the same answers
- The source prompts that triggered it
- Links to the matching model answers, including their date and model
Analyze Brand Behavior
Brand mentions in searches over time
This chart tracks the daily share of fanout searches that name any brand you monitor. Select individual brands to compare your brand with competitors. When brands are selected, a dashed line shows the all-brands average for context. Use the trend to see whether a brand is entering the model’s research process more or less often over time. Apply the model and topic filters to locate where a change is happening.Brand injection
Brand injection measures how often a model adds a tracked brand to its searches when the original question did not name that brand. In other words, it shows which brands the model volunteers while researching a neutral question. Your brand is highlighted so you can compare its injection rate with competitors. Select a bar to review the original prompts and every fanout search issued by their answers. Examples where the model introduced the selected brand appear first.See How Models Rewrite Prompts
Words the model adds
This chart surfaces meaningful words that appear in a fanout search but not in its source prompt. Added words can reveal the product categories, features, comparison terms, and brands a model associates with a question. Select a word to see the prompts and searches behind the count. The drawer also distinguishes cases where a word was already in the prompt from cases where the model injected it.Words the model drops
This chart shows meaningful words that appear in the source prompt but not in its fanout searches. Select a word to inspect examples where the model removed it. Dropped words can reveal language the model treats as unimportant when searching. They do not prove that the model ignored the idea in its final answer; they only describe the search rewrite.How much each model rewrites
The rewrite chart compares two forms of overlap for each model:- From prompt — the share of words in the fanout search that came from the original prompt
- Prompt kept — the share of words in the original prompt that the fanout search retained
Searches per answer
This distribution shows how many searches individual answers issued. Use it to see whether the selected model typically relies on one focused search or fans a prompt out across several research paths. The By model table summarizes search volume, answers with fanouts, searches per answer, average search length, average prompt length, and rewrite overlap for every model with data.Turn Fanout Insights Into Action
- Start with Top fan-out searches and sort by Prompts to identify recurring information needs across your market.
- Filter for your brand and competitors to compare which companies models actively research.
- Review Brand injection to find neutral questions where a competitor enters the research process but your brand does not.
- Use added words to identify vocabulary, product attributes, and comparison angles your content may need to address.
- Use dropped words to spot language that may be too vague, unfamiliar, or disconnected from how models search.
- Open the matching answers and citations before deciding what to change. The fanout explains the search path; the answer shows the outcome.
Gauge can only display fanout text that an AI provider makes available. If a model has no fanout data for your current filters, that does not necessarily mean the model performed no research.
Next Lesson →
Content