Query Fan-Out: How to See the Exact Queries AI Runs Behind Your Keyword (Free Methods Included)

Type one question into Google's AI Mode and Google may run a dozen searches on your behalf. You optimized for the question; the citations went to whoever answered the twelve searches.

That mechanism is called query fan-out, and it explains most cases of "we rank #1 but ChatGPT never cites us." This guide covers how it works, and then goes where the existing coverage doesn't: how to extract the actual fan-out queries yourself, for free, and what to do with them.

How Query Fan-Out Works

Fan-out follows a four-step pipeline, and each step is a distinct place where your content wins or loses.

Step 1: Decomposition. The model reads the prompt, judges whether it's complex enough to need multiple angles, and generates sub-queries covering the facets it predicts the user cares about. A prompt about "best CRM for a small agency" spawns sub-queries about pricing, comparisons, integrations, and reviews.

Step 2: Parallel retrieval. The system runs every sub-query at once against its index or the live web. This is the structural break from classic search: one input, many simultaneous retrievals.

Step 3: Evaluation and extraction. For each sub-query, the model scans top results and extracts the specific passage that answers it. It doesn't read your page the way a person does; it lifts chunks.

Step 4: Synthesis. The extracted chunks get woven into one response with citations. A single answer routinely cites one site for pricing, another for reviews, and a third for specs, sources that would have competed for one ranking slot in classic search.

Losing at step 1 means the model never generated a sub-query your content answers. Losing at step 3 means it did, retrieved your page, and couldn't cleanly extract the answer. Those are different problems with different fixes, and the rest of this guide treats them separately.

How to Extract Real Fan-Out Queries for Free

Every article on this topic tells you fan-out queries exist. Almost none show you how to pull them. Four methods, ranked by usefulness.

Method 1: Pull Fan-Out Queries From the Gemini API

The Gemini API returns the actual search queries it generated inside the grounding metadata of any grounded response. This is the closest thing to reading Google's homework, and it costs fractions of a cent per prompt.

from google import genai
from google.genai import types

client = genai.Client(api_key="YOUR_KEY")

response = client.models.generate_content(
    model="gemini-2.5-flash",  # use the current grounding-capable model
    contents="best crm for a small marketing agency",
    config=types.GenerateContentConfig(
        tools=[types.Tool(google_search=types.GoogleSearch())]
    ),
)

meta = response.candidates[0].grounding_metadata
print(meta.web_search_queries)   # the fan-out queries
for chunk in meta.grounding_chunks:
    print(chunk.web.uri, chunk.web.title)   # who got cited

Two outputs matter here. web_search_queries is the fan-out list, and grounding_chunks shows which URLs won each retrieval, which turns this into a citation-competitor report as well.

Run the same prompt five times and diff the lists, then run your ten most valuable prompts through a loop and export to CSV. That's a fan-out research workflow for the price of a coffee, no enterprise platform involved.

Method 2: Watch ChatGPT's Searches

ChatGPT exposes its search activity in the interface: expand the search step on any response and the queries are listed. The free Keyword Surfer Chrome extension also surfaces them alongside the chat.

This method costs nothing and reflects real behavior, but it's manual. Use it to spot-check your highest-value prompts rather than for bulk research.

Method 3: Read AI Mode's Disclosure

Google AI Mode sometimes announces its work, showing that it's kicking off a specific number of searches for a complex query. You won't always get the full query list, but the count alone tells you how aggressively a topic fans out.

Complex comparative prompts triggering many searches signal high fan-out surface area, meaning more sub-queries you can win. A prompt that triggers no visible search activity is being answered from model memory, and that's a different visibility problem entirely.

Method 4: Ask the Model (with a Caveat)

Prompting "what searches would you run to answer X" produces a plausible facet list, and it's useful for brainstorming. It is not a log of real behavior, so treat the output as hypothesis, never as data.

What Fan-Out Queries Actually Look Like

Fan-out queries cluster into predictable facet types regardless of topic. Pull the queries for any commercial prompt and you'll see the same skeleton.

The individual query strings change between runs. The facet skeleton barely moves, and that distinction drives the entire optimization strategy below.