Knowledge baseSourcesSuganthan Mohanadasan

ChatGPT Rebuilt Its Search Tool. I Read the New Language It Speaks.

Author Suganthan Mohanadasan Date 2026-08-21 Model state GPT-5.6, Plus account, Dubai, DevTools captures 16-20 Aug 2026 Open source →

Key findings

Mohanadasan documents, from dated captures, that OpenAI switched the format of the web.run calls between 16 and 20 August 2026 from JSON objects back to a line-by-line pipe syntax, the third format change since June. He describes the parameters (type, query, window in days, domain), lists call types and windows per topic, and finds that widgets produce no citations. His extension FanoutFox had to be patched because of it.

Why this source matters for the project

It is the first public description of the pipe syntax and the only dating of the switch. It shows what a tracker with daily measurement would deliver: the change would have been on record as a confirmed event on 17 or 18 Aug, not five days later as a reconstruction.

Checked against our data

The article appeared one day after our first full measurement day and explains why we never saw JSON. The dating to 16-20 Aug is not verifiable for us (first run 20 Aug), the syntax itself is attested across all runs. Beyond his list we see slow, search and open as line types and length|medium as a parameter line. The windows match where the topics overlap.

ClaimStatusEvidence
web.run calls have been line-by-line pipe syntax call_type|query|freshness|domain instead of JSON since 16-20 Aug 2026confirmedAll 33 runs from 20 Aug show only pipes, including in thinking mode. Introspection K14 confirms the format independently.
Call types fast, product, business, image, genui_runconfirmedfast, product, business, image observed; additionally slow, search, open, length. genui_run only named via introspection.
Freshness windows in days, topic-dependent (stocks 2, sport 7, commerce 30, financial advice 90, Reddit 3650)confirmedIn our data 1 (news, opening hours), 7 (prices), 30 (local, short search), 90 (product recommendation), 3650 (stable facts: biography K6-01, device specs K7-02). Divergence: 3650 stands for us not for Reddit but for fact stability; 2 (stocks) not observed, as no stock prompt.
Widgets produce zero citationsconfirmedK5-01 thinking with suggest_automation: 0 citations on both days. K1-01 thinking with map_widget: citations come from the entities, not from the widget.
Reddit citations fell from 3.83% to under 1%unverifiableNo Reddit domain in our hits on three measurement days; German prompt set.

Related

mohanadasan-2026-08-10 mohanadasan-2026-07-14 mohanadasan-2026-06-24