Knowledge baseSourcesTomek Rudzki (Peec AI)

Patterns we see in ChatGPT query fanouts

Author Tomek Rudzki (Peec AI) Date 2026-05-05 Model state 5M fan-outs from the Peec panel, several countries, spring 2026 Open source →

Key findings

Statistics over five million fan-outs: count per engine, injected modifiers ('best', year numbers, 'review'), language distribution, reference to RRF as the fusion logic.

Checked against our data

The modifier injection we see daily, in German. The count per prompt is higher for us; the difference may be down to the counting method (semicolon lists, business lines with several queries) or the model generation. The metric fanout-count is the point of connection.

ClaimStatusEvidence
ChatGPT averages 2.1 fan-outs per prompt (Perplexity 1.4, Grok 6.8)outdatedMedian 3 for us, range 2-7 in the standard mode (2026-08-22). Semicolon multi-queries counted, which Peec may not.
'best' is injected in 24.3% of fan-outsconfirmed'beste' in K1-01, K2-01 (fast lines); for German prompts as a German word.
Reciprocal Rank Fusion (k=60) as the fusion method, per YesilyurtunverifiableServer-side. The field ranking_score_present is false for us.

Related

wells-2026-02-12 yesilyurt-rrf