AI Fan-Out Query Anatomy22,518 AI hotel sub-queries decoded
TL;DR: When you ask AI “best hotels in Paris,” it quietly runs 4.5 sub-queries on average before answering. Across 22,518 fan-outs from 5,000 hotel prompts, the modifier “near” appears in 23.2%, “best” in 18.5%, “boutique” in 11.8%. Gemini fans out 7.7× more than ChatGPT — a clue to how each model grounds its answers.
Executive Summary
A fan-out query is what AI types into search before it answers you. Now we have 22,518 of them.
Modern AI assistants don’t answer your question directly. They first generate a handful of sub-queries — fan-outs — that they run against web search, knowledge graphs, or proprietary indexes. The fan-outs determine which sources the AI sees. The fan-outs determine which hotels can possibly be recommended. Every AEO conversation is, ultimately, a conversation about fan-outs.
We pulled 22,518 fan-outs from 5,000 hotel-related prompts across ChatGPT, Gemini, Perplexity, and Grok in April 2026. Three things stand out: AI fans out around location and amenity, not brand; fan-out volume is wildly different across platforms (Gemini is 7.7× ChatGPT); and the AI’s modifier vocabulary is narrower than most SEO keyword lists assume.
How many fan-outs does AI generate per question?
Average: 4.5 fan-outs per capture. But the average hides a highly skewed distribution — the most common count is 2 fan-outs (26.2%), followed by 1 fan-out (18.7%). About 6% of prompts trigger 10 or more fan-outs, with extreme cases hitting 15+. The distribution is bimodal: most prompts get a quick 1-2 fan-out check, but ambiguous or research-heavy prompts trigger long fan-out cascades.
What this means for AEO
If your hotel only ranks for the user’s exact prompt phrasing, you miss every fan-out where AI rephrases it. Optimising for AEO means optimising for the set of fan-outs an AI is likely to generate — not the literal user query.
Modifier vocabulary: what AI actually adds
We checked each of the 22,518 fan-outs for 17 common hotel-search modifiers. Three patterns emerge: location dominates (“near” in 23%), quality framing follows (“best” 18%, “top” 5%), and style/segmentation trails (“boutique” 12%, “luxury” 8%). Notable absences: “business” appeared only 8 times across 22,518 fan-outs — AI does not fan out on the business-traveler axis.
Full modifier table
| Modifier | Mentions | % of fan-outs |
|---|---|---|
| near | 5,226 | 23.2% |
| best | 4,175 | 18.5% |
| with | 3,569 | 15.8% |
| boutique | 2,655 | 11.8% |
| luxury | 1,731 | 7.7% |
| family | 1,693 | 7.5% |
Gemini fans out 7.7× more than ChatGPT
Of the 5,000 captures we sampled, the share that actually expose fan-out queries varies wildly by platform. Gemini exposes fan-outs in 2,478 captures, Grok in 1,745, Perplexity in 456, and ChatGPT in 321. That’s a 7.7× difference between Gemini and ChatGPT.
Why the gap?
Gemini grounds its answers in Google Search and surfaces every sub-query in the API response. ChatGPT often runs a single web_search and synthesizes from training data plus the snippets. The functional impact: Gemini sees a wider but possibly noisier set of sources per answer; ChatGPT relies more heavily on training data plus a narrower web check. For hoteliers, this means a Gemini-shaped strategy and a ChatGPT-shaped strategy differ materially — one optimises for fan-out coverage, the other for training-data presence.
Ten real fan-outs from the corpus
A taste of what AI actually types into search. Notice the density: 2-3 modifiers per query, and several queries that include the year (“2025 2026”) — an AI tic worth its own study.
Methodology
Source
Bright Data’s SERP and AI scraping API exposes a fan_out_queries field for ChatGPT, Gemini, Perplexity and Grok responses. We capture this field as JSON in our fanout_captures table on every weekly scrape.
Sample
5,000 captures from the most recent 60 days, all with a non-null fanout_queries array. Captures are drawn from a 616-prompt library covering 56 city/country combinations and 12 query intents (luxury, family, romantic, business, etc.).
Aggregations
- Volume: count of fan-out items per capture, distribution and average.
- Length: word count per fan-out, computed as whitespace-split tokens.
- Modifiers: 17 hand-picked modifier tokens (best, near, with, boutique, luxury, family, star ratings, romantic, top, affordable, breakfast, rooftop, spa, pool, review, cheap, honeymoon, kids, business). Counted via word-boundary regex.
- Per-platform: count of captures with non-empty fan-out lists, grouped by source.
Limitations
- ChatGPT and Perplexity expose fewer fan-outs not necessarily because they generate fewer — their APIs may not surface the full sub-query trace. The 7.7× gap is an observed-exposure ratio, not a guaranteed generation ratio.
- Modifier token list is hand-picked; long-tail modifiers (e.g., “dog-friendly”, “adults-only”) are under-counted.
- Word counts use simple whitespace splitting; CJK and multi-word entity tokenisation is approximate.
Frequently asked questions
Summarize with AI
