Query fan-out: the sub-queries AI search runs when you ask a question

That description comes from Google Search Central’s documentation on AI features, last updated on 10 December 2025. The name is literal: one query that opens out like a fan into many.

A sub-query is each of the searches the engine runs on its own, starting from the original question. The person types one sentence; the engine turns it into several narrower searches, each about a subtopic, and reads the results of all of them before it writes.

Google introduced the technique alongside AI Mode. In its Google I/O announcement of 20 May 2025, Elizabeth Reid, VP and Head of Google Search, explained that AI Mode breaks the question into subtopics and issues many queries at once on the user’s behalf, so it can go deeper into the web than a traditional search.

Neither sub-query repeats the question as it was written. The first looks for the main fact, the winner; the second looks for a detail that completes the answer, the score of the final.

Applied to a purchase, the mechanism is the same. A question such as “which payroll software suits a small business” can open out into searches about pricing, comparisons between brands or legal requirements. Each of those searches has its own list of candidate pages.

The table stays with Google because these are the surfaces where the vendor itself documents the technique under this name. How other assistants, such as ChatGPT or Perplexity, decide which sources to use is covered in how AI models choose their sources, with what is observable about each one.

The study ran each question 3 times across five sets: 74 tech-hiring questions in English, 73 recruiting-tool questions in Spanish and in English, and 95 AI-visibility questions in Spanish and in English. The three sectors are those of Kulturo, HeyTalent and DoGeo itself.

“No web search” means that run issued no sub-query at all, not that the answer was poor or made up. The percentages hold for that model, that week and those sets: they do not describe Gemini in general, or any other assistant.

The gap appears in both sectors measured in both languages, and it points the same way. The study measures what the model does, not why, so it does not support pinning the gap on a specific cause.

For a company selling in Spanish, the practical consequence is that a larger share of Gemini’s answers is written from memory. There, the company’s own page plays no part at the moment of answering: what counts is what other sources published about the brand before training.

Ten sources for two searches means each sub-query feeds several pages into the answer. A page that ranks well for just one of those searches can already make the source list, even if it does not appear for the other.

The year gets added to sub-queries only on some topics. In the same study, between 11% and 15% of the recruiting sub-queries contained “2025” or “2026”; on the GEO questions, 1%. Whether the model adds the year depends on the topic; it is not a fixed rule.

• Being in Google’s index is the entry condition. According to Google Search Central, a page can only appear as a supporting link in AI Overviews or AI Mode if it is indexed and eligible to be shown in Google Search with a snippet; there are no additional technical requirements.
• Every page has to answer one subtopic on its own. A sub-query about pricing finds the pricing page, not the home page; one about a comparison finds the comparison. How to track whether assistants mention a company is covered in the guide to rank tracking and AI visibility.
• What assistants cannot read does not compete. Before thinking about sub-queries, it pays to check whether AI crawlers can read your site, because a block at the CDN shuts every page out at once.
• When the model does not search, third-party sources count. With questions in Spanish, the DoGeo study found more answers without a search, and there what other sites say about the brand carries the weight.

The concrete steps for working on presence in this assistant are in the guide to appearing in Gemini.

Bot logging works with a rule that records, without blocking, the requests from each bot, identified by its user-agent (the name it announces itself with) and verified by IP when the provider publishes its ranges. Those logs show which pages each bot visits and when. This is what DoGeo sets up for each client.

Each bot tells you something different. OAI-SearchBot, according to OpenAI, is used to surface websites in ChatGPT’s search results, so its visits show which pages enter that index. ChatGPT-User and Claude-User open a page when a person asks a question, according to OpenAI and Anthropic, and they are the closest signal that the assistant has read the page in a conversation.

The log does not see what the assistant writes in its answer, nor mentions that need no visit to the site. For that, running a fixed set of questions and noting who appears is still necessary.