LLMRanks · Original Research
00 / 09 · Query Fan-Out
Original research · June 2026 · How AI search really works

You ask AI one question.
It secretly runs a dozen.

When you ask ChatGPT, Gemini, or Google a question, it doesn't search what you typed — it quietly fans out into a whole list of background searches you never see. We captured the real ones. They decide whether your business gets mentioned at all.

3 buyer queries 2 engines, real searches 1 that hides them June 2026
01
The mechanism

One question becomes
a fan of searches.

AI search doesn't answer from a single lookup. It decomposes your question into multiple background searches, retrieves sources for each, and synthesizes one answer. Google calls it "query fan-out." The consequence: the thing that gets cited is a passage answering a sub-query — not a page ranking for your keyword.

Query fan-out diagram: one question fans out into many hidden background searches, which converge into one synthesized answer
One question → many hidden searches → one synthesized answer

Google, in its own docs: AI Overviews & AI Mode use a "query fan-out technique — issuing multiple related searches across subtopics and data sources." Google's patent (US11663201B2) calls the underlying technique "query variants." "Fan-out" is the industry's name for it.

02
The reveal

So here are the
searches it actually ran.

One simple buyer question. These are not our guesses — they're the real sub-queries each engine fired, captured straight from the engines.

You asked "best robot vacuum for pet hair"
ChatGPT3 searches
best robot vacuum pet hair 2026 reviews
Wirecutter best robot vacuum pet hair 2026hunts an authority site
RTINGS best robot vacuum pet hair robot 2026hunts an authority site
Gemini8 searches
"Ecovacs Deebot X8 Pro Omni" pet hair reviewspecific model
"Roborock S8 MaxV Ultra" pet hair reviewspecific model
"MOVA P50 Pro Ultra" pet hairspecific model
"Roborock Saros 20" pet hairspecific model
best robot vacuum for pet hair 2026
best robot vacuum for pet hair 2025
rtings best robot vacuum for pet hair
wirecutter best robot vacuum for pet hair
In LLMRanks · Generate we pull this real fan-out for your topic and write the article to answer every one of these searches.
03
The shortlist

The AI didn't know these
vacuums. It found them.

You named no brand. So where did the specific models come from? We ran the same question two ways — memory only, then with web search on — and watched what changed.

The question "best robot vacuum for pet hair"
① Web OFF — from memoryno search
Roomba j9+
Roborock S8 Pro Ultra
Roomba j7+
Shark Matrix Plus
Eufy X10 Pro Omni
It just recalls — all 2023–24 models · cutoff ~Jan 2025
② Web ON — what it searchedlive
best robot vacuum for pet hair 2026
wirecutter best robot vacuum for pet hairchecks a trusted source
rtings best robot vacuum for pet hairchecks a trusted source
"Roborock Saros 20" pet hairnewer than its memory
"MOVA P50 Pro Ultra" pet hairnewer than its memory
It doesn't trust memory — it searches, and finds models it never knew

Same question — memory gives you 2024's answer; search gives you today's, found live.

How it gets there — a two-stage fan-out
1 · Search the trusted review sites first
Wirecutter · RTINGS · Forbes Vetted
2 · Read who they rank right now
today's winners, not 2024's
3 · Fire a second wave to vet those models
the specific-model searches
The takeaway for LLMRanks · Get cited — AI's shortlist is whatever the review sites rank today, not what the model memorized. You get on it by getting into those sources.
04
The pattern

There's no magic number —
it scales with the question.

Change the shape of the query and the fan-out changes with it. A comparison fans out wide; a narrow how-to collapses to almost nothing. Anyone quoting a fixed "5 to 30 searches" is guessing — here's what really happened.

The questionChatGPTGemini
best robot vacuum for pet hair buyer38
Roomba vs Roborock comparison65
how to choose … small apartment how-to13
ChatGPT → comparison6 searches
Roomba vs Roborock 2026 robot vacuum comparison
site:us.roborock.com Saros 10R Roborock officialgoes straight to the official site
site:irobot.com Roomba 505 Combo robot officialgoes straight to the official site
Gemini → comparison5 searches
"iRobot" bankruptcy 2025 2026pulls live news
iRobot Picea acquisitionpulls live news
Roomba vs Roborock 2025 2026

A practitioner study (Seer, Nov 2025) measured ~10.7 fan-out searches on average for Gemini 3 — but Google itself publishes no number. Treat any count as "it depends on the question," never a fixed figure.

05
The hold-out

Google won't show its searches —
but the fingerprints do.

Run the same query against Google AI Mode and scan the entire response: zero sub-queries. Google keeps its fan-out server-side. But you can still read it — because the searches show up in what it cites and names.

Sub-queries exposed — same query, three engines

ChatGPT
3
Gemini
8
Google AI Mode
0
"best robot vacuum for pet hair" · full-payload scan of Google AI Mode = 0 sub-query fields.
But the fan-out is hiding in plain sight
Google's answer leads with the Ecovacs Deebot X8 Pro Omni
the exact model Gemini searched
Google cites RTINGS among its sources
the exact site ChatGPT targeted

Google AI Mode & AI Overviews run on Gemini and use the same documented fan-out — Google just doesn't expose the searches. So the visible Gemini fan-out is your closest window into what Google is doing under the hood.

06
The stakes

This is why ranking #1
no longer saves you.

Because AI cites the passage that answers a sub-query — not the page that ranks for the head term — being #1 stopped guaranteeing the citation. The link between rank and citation has come apart.

38%
of AI citations come from the organic top 10 — down from ~76% a year earlier
31%
come from pages ranking beyond position 100
95%
of fan-out searches have zero traditional search volume
In LLMRanks · Optimize we score a page on the sub-queries it covers — not the head term it ranks for.

Rank-decoupling: Ahrefs, Mar 2026 (863K SERPs / 4M AI Overview URLs). Zero-volume share: Seer, Nov 2025. A keyword tool can't show you 95% of these searches — only the fan-out can.

07
The play

Optimize for the searches —
not the keyword.

The fan-out is a to-do list. Cover the recurring themes as self-contained passages an engine can lift whole. For robot vacuums, the data says:

Answer each sub-theme as its own passagepet hair · mapping/lidar · self-empty · mopping · price tiers · small-space
coverage
Name specific current models & the current yearthe engines search exact model names + "2026" — generic pages don't match
specificity
!
Coverage is necessary — not always sufficientfor the hottest buyer query you're up against Wirecutter & RTINGS; win the comparison & how-to sub-queries, be the expert source, earn the citation
be honest

Don't chase the ephemeral one-offs (a single model that'll be replaced, a news event) — cover the durable sub-query themes. SEO gets you eligible; covering the fan-out gets you chosen.

08
The product

Here's how we build
for the fan-out.

Most tools optimize for the keyword you typed. LLMRanks pulls the real fan-out for your topic and writes the article to answer every sub-query in it — at generation time, before you ever wait on rankings.

See the real fan-out for your topic

Enter a topic → LLMRanks captures the real fan-out → generates an article that answers every sub-query.
09
Your move

Stop optimizing for one question.
Optimize for the dozen AI asks.

You've seen the hidden searches. Now build content that answers them — and see the real fan-out for your own topics.

Build for the fan-out → llmranks.io
llmranks.io