Knowledge baseSourcesSuganthan Mohanadasan

How ChatGPT Actually Picks Sources (I Read the Network Traffic, Not the Outputs)

Author Suganthan Mohanadasan Date 2026-06-24 Model state GPT-5.x, Plus account, Dubai, ~1,240 hit records from DevTools Open source →

Key findings

The first article of the series: network traffic instead of output, discovery of the retrieval pipes via result_source, intent classification turn_use_case, fan-out volume in thinking mode, three-stage funnel.

Checked against our data

Methodically the closest relative of GPTSpy. The fields that carry the article vanished four weeks later, which shows the value of a daily measurement. The fan-out counts in thinking mode are an order of magnitude lower for us; either model state or a difference between "thinking" (level medium) and a higher effort.

ClaimStatusEvidence
result_source per hit: labrador, bright, oxylabs, serpoutdatedField gone since 2026-07-21 (his own follow-up). Never seen in our data.
turn_use_case classifies intent into six buckets; for text no searchunverifiableField not observed.
Thinking model 15-40 fan-outs including site:outdatedThinking is 2-7 fan-outs in our data (K2-01 thinking 7); site: only at K5-01 thinking.
Three stages fetched / cited / mentioned; Reddit 278x fetched, 11x cited; local_results_limit 2confirmedThe funnel is traceable in our data as listed -> opened -> footnote; Reddit does not appear for German prompts.

Related

mohanadasan-2026-07-14 goodwin-2026-07-08