Explore a question's subtopics
Generate labelled query approximations around a main question for research and planning.
Query Fan-Out Simulator
Free, no signup. Map the research questions around a query, label how they were generated, then check lexical coverage against your content.
Separate pages with ---SOURCE---. Start a pasted block with SOURCE: label, or use a bare URL block. URL-only blocks are fetched; pasted content stays in the browser.
Use your own Gemini API key (optional)
The key is sent directly from this browser to Google and is never sent to this site, stored, or logged here. Google may apply project quota or charges. Clear the field when finished.
Requests are processed by our server and are not stored after processing. Generic sub-queries and pasted-source competition run locally. The query may be sent to the selected AI path for optional calibration; URL-only sources are fetched through the protected text endpoint. Pasted source text is not sent for calibration. Anonymous run-level outcome counters may be used for aggregate research; URLs, domains, IPs, and identifiers are never included, and no statistic is released below 100 runs.
Sample report Deterministic generic example
For canonical tags with no page text, the local generator always returns five unique rows:
canonical tags · generic · entity-expansion · mixed
What is canonical tags? · generic · definitional · mixed
canonical tags benefits and limitations · generic · entity-expansion · mixed
canonical tags vs alternatives · generic · comparative · mixed
How to choose canonical tags · generic · implicit · mixed
Because no source text was supplied, every row says to paste page content rather than inventing a coverage percentage.
How to use it
- Enter the main query you want to explore.
- Optionally paste a page’s text or HTML to add a lexical coverage check.
- Optionally enter your Gemini API key for a direct Google request. Otherwise the tool tries the bounded site model and falls back to the deterministic generic set.
- Select Simulate fan-out, keep only relevant sub-queries, then review gaps against the strongest page chunk before editing.
Site passport Local context for this saved site
Local data
Saved targets, named lists, and recent check summaries remain only in this browser.
Rate this tool
What the simulation shows
- generic — a visible deterministic template, not provider behavior.
- calibrated — a model-generated approximation for this run, labeled as Gemini or Workers AI provenance.
- Type — a rule-based intent label such as comparative, freshness, or definitional.
- Answer mode — a heuristic guess that the task is memory-, retrieval-, tool-, or mixed-led.
- Covered / Gap — whether the strongest lexical chunk reaches the 55% query-term coverage threshold.
How it works
The fallback expands the entered query with four fixed research shapes, removes duplicates, and classifies each row from visible keywords. With a key, the browser sends a query-only prompt directly to Gemini 3.5 Flash and parses one sub-query per line. If that fails, the bounded Workers AI path is attempted before the generic fallback. For coverage, supplied content is extracted, chunked at roughly 300 tokens with overlap, ranked lexically, and compared with a 55% threshold. Generated rows are approximations; no provider’s private query logs are accessed.
Features
- Always-available deterministic fan-out.
- Optional direct BYO-key Gemini approximation without routing the key through this site.
- Multi-page, per-sub-query lexical retrieval competition.
- Winning page and chunk with visible matching terms and runner-up coverage.
- Consensus versus corpus-unique claims, numbers, and quotes with explicit originality and support caveats.
Limitations
Real answer systems can plan, search, and reformulate in ways this simulator cannot observe. Generic wording can be awkward, classifications are keyword rules, and lexical matching misses synonyms and deeper semantic equivalence. A generated row can also be irrelevant to the page’s intent. Treat the map as an editorial brainstorming aid, not evidence of ranking or inclusion in an AI answer.
AI answers may continue to reflect cached or previously retrieved material for days to weeks after a fix. Recheck over time before treating an unchanged answer as proof that the fix failed.
Frequently asked questions
What is query fan-out?
Query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics. is the use of multiple related searches or retrieval tasks to help answer a broader prompt. Search providers may use it, but the exact sub-queries for a live answer are generally not exposed.
Are these the queries Google or an AI system actually used?
No. Generic rows are deterministic templates and calibrated rows are a model-generated approximation. The tool labels provenance explicitly and never presents either tier as observed product behavior.
How is page coverage calculated?
The page is split into chunks and ranked locally with lexical retrieval. Coverage is the share of query terms represented in the strongest chunk; a row is marked covered at 55 percent or higher.
Does a coverage gap mean I should add a new section?
Not always. Add material only when the sub-query is relevant to the page’s purpose and you can answer it accurately. Memory-led, off-intent, or redundant rows can be ignored.
Does the tool use embeddings?
No. The current coverage check uses BM25/cosine-style lexical components and reports that embeddings were not evaluatedEmbeddings are dense numerical vectors — lists of floating-point numbers — that represent the meaning of text in a high-dimensional space. Semantically similar content lands close together, so search and AI systems can match by meaning, not just keywords.. Semantic coverage can therefore be missed when a passage uses different wording.
What happens to my Gemini API key?
It stays in the password field for this tab and is sent directly from your browser to Google in the x-goog-api-key request header. The site does not receive, store, log, or persist it. Google API usage can consume your project quota or incur charges under your Google account.
Feature requests for Query Fan Out
Upvote what you want most. New ideas can be submitted from the floating Feedback menu; requests appear here once approved, and the most-wanted rise to the top.
You won't be emailed about that request anymore.
Loading…
➕ Request a feature
New requests are reviewed before they appear here.
Where this tool helps
Common use cases
Check page coverage
Compare the supplied page with generated sub-queries to find concepts it addresses weakly or not at all.
Build a content brief
Group relevant fan-out ideas into sections, FAQs, comparisons, or supporting pages for editorial review.
Avoid overclaiming private behavior
Use the simulation as a planning aid without presenting it as observed queries from a proprietary search system.
Watch the full workflow
Query Fan-Out Simulator walkthrough
Read the transcript
Query Fan-Out Simulator
A broad question may involve several smaller research tasks, but real product sub-queries are usually hidden. I’ll show you how to generate an explicitly labeled approximation, compare multiple supplied pages, inspect lexical winners and gaps, interpret information gain carefully, understand privacy and limitations, and turn relevant rows into editorial next steps.
Step 1
Query fan-out means using related searches or retrieval tasks around a broader prompt. This tool makes those research shapes visible, then checks supplied pages lexically. Generic and calibrated rows are approximations. Neither is evidence of the exact sub-queries used by Google or another live product.
Step 2
Use it while planning a guide, comparing two owned pages, consolidating overlapping content, or reviewing whether an important question is answered directly. It can reveal useful definitional, comparative, decision, and limitation angles, but every row still needs an intent check.
Step 3
Enter the main query and optionally add page blocks separated by dash-dash-dash Source. Pasted content stays local for competition. Select Simulate fan-out and competition. The tool tries a bounded calibration path, then honestly falls back to the deterministic generic set when the model is unavailable.
Step 4
Actual execution surface records whether a model ran, which route was requested and resolved, the query time, and retrieval state. A not-evaluated model with generic rows is still a valid deterministic workflow. Never relabel those rows as observed provider behavior.
Step 5
The summary states how many sub-queries and supplied sources were analyzed. Competition uses lexical B-M-twenty-five and cosine-style components only. Embeddings, authority, factuality, live rank, and a provider retrieval trace are explicitly not evaluated.
Step 6
Each row shows the sub-query, generic or calibrated tier, rule-based intent type, and a heuristic answer mode such as memory, retrieval, tool, or mixed. These labels organize review; they do not prove how an answer system planned the task.
Step 7
For a positive match, the row names the winning page, chunk, query-term coverage, excerpt, matching terms, and runner-up coverage. Open the excerpt and judge whether it truly answers the sub-query. Term overlap can reward shallow or misleading text.
Step 8
Covered means the strongest first-source chunk reached fifty-five percent query-term coverage. A gap can be real, irrelevant, or merely phrased with synonyms the lexical method missed. Check the page purpose and semantic meaning before deciding to add anything.
Step 9
Information gain comparison surfaces consensus-like text and claims, numbers, or quotes unique within this small supplied corpus. Unique does not mean original, true, valuable, or absent from the web. Partial sources also prevent reliable absence and uniqueness conclusions.
Step 10
The optional Gemini key stays in this tab and is sent directly from the browser to Google, where quota or charges may apply. Clear it afterward. Without a key, the site may try its bounded model. Pasted source text remains local for competition, while U-R-L-only blocks use protected fetching.
Step 11
Real systems may plan, search, and reformulate in unobservable ways. Generic wording can be awkward, intent classifications are keyword rules, lexical matching misses synonyms and deeper equivalence, and generated rows can be off-intent. This is brainstorming support, not ranking or citation evidence.
Step 12
Download the C-S-V, keep only sub-queries aligned with the page, and review the winning chunks. Research genuine gaps, add accurate material without duplicating another page’s purpose, rerun the same sources, and compare results with real query data and human editorial judgment.
Keep only the questions that serve the page.
Export the useful rows, discard awkward or off-intent suggestions, and investigate each real gap in the strongest source chunk. Add material only when you can answer it accurately and it belongs on the page. Then rerun the same corpus and validate changes with real audience and search evidence.
Common issues & how to fix them
- Warning Page leaves a sub-query uncovered Fix: Add a self-contained passage that directly answers the uncovered sub-query in the answer format its intent requires.
- Information Page only partially covers a sub-query Fix: Expand the best matching passage with the missing criteria, steps, entities, or evidence needed to answer the full sub-query.
- Warning Best matching passage is weak Fix: Rewrite the matching passage with a direct answer, precise query terms, and supporting evidence that stands on its own.
- Information Fan-out branches overlap Fix: Merge overlapping branches into one distinct question or rewrite them so each branch covers a separate intent.