Knowledge baseSourcesDavid Konitzny (Peec AI)

How ChatGPT Deep Research reads your site: What the logs reveal

Author David Konitzny (Peec AI) Date 2026-06-22 Model state Deep Research, 10+ free accounts, ~20 requests per account, WebSocket logs validated against HTML crawls Open source →

Key findings

Konitzny logs by WebSocket how Deep Research reads pages: search returns Bing snippets of about 250 characters, open loads a window of median 5,700 characters, find searches within it like Ctrl-F. The reader is a text browser without JavaScript that never clicks. The first hit gets 40% of all openings; pages with many navigation links are read as content to about a third. Alt texts are read, being read does not mean being cited.

Why this source matters for the project

Methodically the closest counterpart to GPTSpy in the DACH region: network capture instead of output statistics, several accounts, validation against own crawls. The surface split (standard, thinking, deep research) is the axis along which our numbers and his can be laid side by side.

Checked against our data

The article describes Deep Research, not the standard or thinking mode, and relies on free accounts and Bing. Its snippet length fits neither of our two signatures, which supports the separation of the pipelines by surface. The operations open and find match the names our introspection prompts (K14) return. His older observation that German users are researched in English we see confirmed for entity and image questions.

ClaimStatusEvidence
Three steps Search -> Open -> Find; Find is a text search inside the opened documentunverifiableDeep Research not in the prompt set. open observed (Thinking), find only named via introspection K14.
Snippets ~250 characters, max ~285 (Bing)outdatedIn our data 202 (index) and 131-158 (scrape). Deep Research uses Bing per the article; our modes do not.
Reading window ~5,700 characters median, text browser without JS, never clicksunverifiableServer-side.
First source gets 40.7% of all openings; pages with 60+ links only 33% read as contentunverifiableNo open in the standard mode.
German prompts are researched internally in English (podcast 2025)confirmedEnglish fan-outs at K6-01 (OpenAI leadership Sam Altman CEO biography), K11-01 (Allianz Arena Munich at night); mixed at K2-01. Metric from schema 14.

Related

konitzny-2025-11-19