Query fan-out `term/query-fan-out`

Terms

What it is

From one user question ChatGPT builds several search queries that go to the retrieval layer in parallel. That is the fan-out. We see it as lines in the code message to web.run, one line per query, with mode, text, optional time window and domain. Several queries in one line are separated by semicolons.

Measured 22 Aug 2026: two to six fan-outs per searching question in standard mode, median three. Thinking mode on product questions up to seven. Wording differs every day; the count varies by about one. Example K1-01: business|Munich, Bavaria, Germany|Laufschuhe Laufanalyse Laufberatung Running Shop;Laufschuhberatung Laufbandanalyse plus fast|beste Laufschuhberatung München Laufanalyse Running Shop|30.

Notable: fan-outs often carry things the user never said. "best", the year 2026, "official", "Germany", concrete product names, in K8-01 even a phone number. The model knows what it is looking for before it looks.

External evidence

Tomek Rudzki, Peec, 5 May 2026: 5M fan-outs, ChatGPT 2.1 per prompt, "best" injected in 24.3 %.
Tom Wells, Peec, 12 Feb 2026: fan-out length doubled from 6 to 12 words (Oct 2025 to Jan 2026), count stable at 2.3–2.8.
Josh Blyskal, 8 Dec 2025: 36.4 % of prompts trigger two, 52.9 % three queries; ChatGPT appends a year in 17 % of fan-outs (22 Jul 2026).
Suganthan Mohanadasan, 10 Aug 2026: brands in the fan-out are cited 33× more often than brands only retrieved.

For visibility

Whoever is absent from the fan-outs is never checked. That holds for brands in product fan-outs as much as for terms: a shop without "Laufanalyse" (gait analysis) on its page does not appear for K1-01, however good its advice. The fan-outs are the real keyword set for ChatGPT, and they differ from what people type into Google: longer, with modifiers, partly English, partly with a year.

External sources

mohanadasan-2026-08-10 · rudzki-2026-05-05 · wells-2026-02-12 · blyskal-2025-12-08

Change history

No event has touched this object yet.

Open

Share of English fan-outs for German prompts. Wells/Peec: 78 % of non-English prompts carry at least one English fan-out. Visible in our data (K6-01, K11-01), not yet counted — metric fanout-language-en-share from schema 14.
Why do product fan-outs already name concrete models (K2-01: Sage Bambino Plus, Gaggia Classic E24, DeLonghi Dedica) before any search? Mohanadasan: 68.9 % citation rate for brands named in the fan-out vs 2.1 % for brands merely retrieved.

web.run syntax (pipe format) Freshness window Domain restriction Search decision Fan-outs per question