Query fan-out › Query fan-out examples

Query fan-out examples

From QueryFanout.wiki, the technical reference and research archive for query fan-out

This page shows what a query fan-out looks like. Every example carries one of three labels, and the labels are never mixed in one table. OBSERVED means the platform displayed the subqueries and we captured them as displayed. RECONSTRUCTED means the subqueries were not shown and have been inferred from what was retrieved or cited. SYNTHETIC means the decomposition was written as an illustration and was never produced by any search system.

Google AI Mode does not display its subqueries, and in the current desktop Chrome build AI Mode opens inside a browser-internal page that automation cannot read (see Measurement). There is therefore no OBSERVED example from Google on this page. Every observed example below is from Perplexity, which shows its searches on request. Do not read Perplexity's behaviour as Google's.

Contents

Google's own illustration
Observed fan-outs
The same prompt three times
Prompts that did not fan out
A follow-up turn
A reconstructed fan-out
Synthetic decompositions
Patterns worth noting
References

Google's own illustration

DOCUMENTED The only worked example we located in Google's public documentation as of September 2026. It is an illustration written by Google ("might include"), not a capture.

how to fix a lawn that's full of weeds

best herbicides for lawns
remove weeds without chemicals
how to prevent weeds in lawn

Google Search Central, "Optimizing your website for generative AI features on Google Search", May 2026.^[1] Three subqueries: a product angle, a method angle, a prevention angle.

Observed fan-outs

OBSERVED All from Perplexity, signed-in free account in Colorado, 8 September 2026 UTC (the evening of 7 September local time), captured from the platform's own "steps" panel. Full records with source counts and cited domains are in the dataset.^[2]

Perplexity answer for the sourdough prompt with the Finished 1 step panel expanded, showing the step label Checking how to avoid drying layers on sourdough starter, the three searches it ran, three of the pages it read, and the first paragraph of the answer — What an observed fan-out looks like on Perplexity: the step label, then the searches it ran, then the pages it read. Rendered from the page as displayed on 8 September 2026 UTC, record qfo-0003; site icons are omitted by the capture method.

Long-tail informational (qfo-0003)

How do I stop my sourdough starter from developing a grey liquid layer on top

sourdough starter grey liquid layer on top what is it
how to prevent hooch sourdough starter grey liquid
sourdough starter gray liquid hooch fix

Step label: "Checking how to avoid drying layers on sourdough starter". 16 sources retrieved, 4 domains cited. The prompt never says "hooch"; the system introduced the term before searching, and two of the three subqueries depend on it.

Comparison (qfo-0004)

Sony WH-1000XM6 vs Bose QuietComfort Ultra for long flights

Sony WH-1000XM6 vs Bose QuietComfort Ultra long flights
Sony WH-1000XM6 battery life ANC comfort review 2025 2026
Bose QuietComfort Ultra battery life ANC comfort long flight review

Step label: "Searching for the latest long-flight reviews". 15 sources retrieved, 9 domains cited. The pattern is one comparison query plus one query per entity, each with the same attribute list (battery, ANC, comfort) and a year token added.

Commercial recommendation (qfo-0005)

Best standing desk under $400 for a small apartment

best standing desk under $400 2025 2026
compact standing desk small apartment under $400
standing desk reviews small space budget

Two steps. Step 1, "Searching for apps under $400", ran the three searches above. Step 2, "Looking up current small-space standers", was not a search: it fetched three specific pages by URL (finforum, cnet, thesweetandsimplekitchen). Fetches are recorded separately from subqueries. 15 sources, 3 domains cited.

Rare topic (qfo-0007)

Traditional dairy dishes served at Tsagaan Sar in Uvs province Mongolia

Tsagaan Sar traditional dairy dishes Uvs province Mongolia
Mongolian Tsagaan Sar food dairy products Uvs
traditional Mongolian dairy dishes holiday Tsagaan Sar

Two subqueries keep the rare constraint (Uvs); the third drops it and generalises to the national holiday. The answer then admits that Uvs-specific dishes "are not widely documented in public sources". The fan-out hedged against an empty result.

Freshness-sensitive (qfo-0008)

What changed in the most recent iOS update

iOS 26.0 update changes features
Apple iOS latest update September 2026 what's new
iOS 26 release notes changes

23 sources retrieved, the most in the batch. Two of the three subqueries name iOS 26. The answer says the most recent major update is iOS 27. The decomposition guessed a version and the synthesis corrected it from what came back. This is the clearest example in the batch of fan-out and synthesis disagreeing.

Local (qfo-0009)

Best time to visit Great Sand Dunes National Park to avoid crowds

best time to visit Great Sand Dunes National Park avoid crowds
Great Sand Dunes National Park least crowded months weekdays
Great Sand Dunes National Park crowd patterns seasonal tips

A near-paraphrase of the prompt plus two angle queries. 15 sources, 3 domains cited.

Product (qfo-0010)

Is the DJI Mini 5 Pro worth it over the Mini 4 Pro

DJI Mini 5 Pro vs Mini 4 Pro differences specs
DJI Mini 5 Pro worth it over Mini 4 Pro review
DJI Mini 5 Pro release date price features

Step label: "Comparing features and prices". Specs, verdict, and price-and-date angles. 17 sources, 6 domains cited.

The same prompt three times

OBSERVED One prompt, three fresh threads, twelve minutes apart, 8 September 2026 UTC. This is the repeat-run test described in the methodology.

	Run 1 (03:08 UTC)	Run 2 (03:19 UTC)	Run 3 (03:20 UTC)
Prompt	What is the best CRM for a 50-person law firm concerned about security and migration cost
Step label	Researching secure and cost-effective CRM options	Searching for secure CRM options for a small law firm	Comparing cybersecurity considerations
Subquery 1	best CRM for law firm 50 employees security	best CRM for law firm 50 employees security	best CRM for law firm 50 employees security
Subquery 2	legal CRM migration cost considerations 2025 2026	legal CRM migration cost considerations	legal CRM migration cost comparison 2025 2026
Subquery 3	top secure CRM platforms for mid-size law firms	top CRM for mid-size law firms 2025 2026 security compliance	law firm CRM security compliance Clio Lawmatics Filevine
Sources retrieved	15	15	15
Domains cited	9	9	11
Top recommendations	Lawmatics; Clio Grow + Manage; Salesforce	Clio; PracticePanther; CosmoLex; Salesforce; Dynamics 365	Clio Grow + Manage; Lawmatics; Filevine

What stayed the same: the first subquery, word for word, all three times. The second subquery kept its meaning and varied its tokens. What changed: the third subquery each time, and almost the entire set of cited domains. Only one domain (spellbook) was cited in all three runs. The recommended products differed in every run. Three runs is a tiny sample, and it is enough to show that stable subqueries do not produce stable answers.

Three columns, one per run of the law-firm CRM prompt, each listing three subqueries and the cited domains; the first subquery is highlighted as identical in all three runs and spellbook is highlighted as the only domain cited in all three — Three runs, twelve minutes apart. Subquery 1 identical every time; the cited domains almost entirely different.

Prompts that did not fan out

OBSERVED Two of the eleven launch prompts produced a single search identical to the prompt. On this platform, at least, fan-out is conditional.

How tall is the Eiffel Tower

How tall is the Eiffel Tower

Step label: "Searching the web", 1 second. 9 sources, 3 domains cited. Simple factual prompt.

jaguar speed

jaguar speed

Step label: "Searching the web", 1 second. 9 sources, 2 domains cited. The ambiguity (the animal or the car) was handled in the answer, which covered both senses, rather than by searching for each sense. Fan-out would have been the obvious way to disambiguate. It did not happen.

A follow-up turn

OBSERVED A second turn in the headphones thread: "Which one has better battery life". The platform showed no step panel at all for this turn, yet attached 9 sources, two of them from domains not present in the first turn. Retrieval happened; the subqueries were not disclosed. Record qfo-0011 is therefore fanout_observable: false, and it is the reason the dataset has that field.

A reconstructed fan-out

RECONSTRUCTED When a platform does not show its subqueries, the only way to say anything about them is to work backwards from what it cited. Here is what that looks like for the follow-up turn above, and it is inference, not observation.

Cited domain	Content it supplied	Information need it answers
headphonecurve	Rated battery figures for both models	manufacturer battery ratings
soundguys, mute-zone	Independent test results for the Sony	Sony WH-1000XM6 tested battery life
techtimes	Real-world figures for the Bose with spatial audio on	Bose QC Ultra battery life with Immersive Audio

A reconstruction can tell you what information needs the response covered. It cannot tell you how many searches ran, what they were worded, or whether the platform ran searches whose results it discarded. That is why a reconstructed row never gets a subquery_count.

Synthetic decompositions

SYNTHETIC Written by hand or generated by a simulation tool. Useful for planning content. Never evidence of what any system did.

What is the best CRM for a 50-person law firm concerned about security and migration cost?

CRM for law firms
CRM security compliance
CRM pricing for 50 users
CRM migration costs
Salesforce vs HubSpot for legal firms
legal CRM integrations
CRM onboarding time

Illustrative decomposition. Compare with the observed runs of the same prompt: the platform ran three subqueries, not seven, and its subqueries were longer and carried more of the prompt's constraints (50 employees, security, migration) in each one.

Simulation tools such as Qforia, queryfanout.io and the LLMrefs generator produce lists of this kind, typically ten to thirty items, by asking a language model what a search system might search for.^[3] They are labelled SYNTHETIC here regardless of how they describe themselves. The Measurement page lists which tools observe and which simulate.

Patterns worth noting

TESTED From eleven prompts and two repeats on one platform. Small sample; patterns, not laws.

Three is the default. Every prompt that fanned out produced exactly three subqueries. Neither more complex prompts (the CRM prompt) nor rarer ones (Uvs) changed the count.
The first subquery is a paraphrase. In nine of nine fan-outs, subquery 1 restated the prompt in search-engine phrasing. It was also the stable one across repeat runs.
Year tokens are added. "2025 2026" appeared in five of the nine fan-outs, never in the prompt.
Comparisons split by entity. One query per product, each with the same attribute list.
The system names things before it searches. "Hooch" for the sourdough prompt; "Clio Lawmatics Filevine" in the third CRM run. The decomposition model brings its own knowledge into the subqueries.
Fan-out size did not predict source count. Single-search prompts retrieved 9 sources; three-subquery prompts retrieved 15 to 23.
Follow-ups hide their searches. The second turn attached sources without showing any step.

References

Google Search Central. "Optimizing your website for generative AI features on Google Search." 15 May 2026, updated 10 July 2026. developers.google.com.
QueryFanout.wiki. "Query Fan-Out Observatory, batch 1." 8 September 2026. Dataset version 0.1.0. queryfanout.wiki/data/. Each record links to the Perplexity thread it was captured from.
Anconitano, Veruska. "Query fan-out tools and software." Search Engine Land, 22 April 2026. searchengineland.com.

Categories:

Information retrieval
AI search
Google AI Mode