Skip to main content
ChatGPT Deep Research narrates its whole session over a recordable WebSocket, and Peec recorded it across 10+ accounts and ~200 runs (full study, Jun 2026). The mechanics (this covers Deep Research’s text browser, not quick Search or Agent mode):
  • Three commands, never a click. search (Bing), open (skim a page as numbered plaintext), find (Ctrl+F inside the open page). click was used zero times across every logged session — content reachable only via JavaScript, accordions, tabs, or filters does not exist for the agent. Client-rendered SPAs are hit hardest.
  • All retrieval runs on Bing (web_with_bing), no Google fallback. Bing snippets cap at ~285 characters — the entire window you get to earn the open. For ChatGPT, only the top 2–3 Bing results per query make the read set once the agent is confident (its topn depth parameter narrows from 10 to 2–3 as it converges).
  • The read is your HTML, linearized top-to-bottom in source order, <head> stripped, nothing JS-rendered. It follows links — including internal ones — by reading the href from inline link markers, so descriptive anchor text routes it to pricing and spec pages.
  • Blocked means invisible. robots.txt blocks on OAI-SearchBot return zero content (this is a different bot from GPTBot, which governs training — see AI bot types). Login walls, bot-refusing marketplaces, and timeouts silently drop sources.
  • Re-reads are keyword-driven: the top trigger is a successful find — the agent re-opens the page positioned on the matched line (95% of the time in the sample). Use the exact words people search, verbatim, in headings and copy.
How much of your page fits in each read, and why navigation costs you: Deep Research reading budget. Exact figures are a June 2026 snapshot; the mechanics are the durable finding. Source: How ChatGPT Deep Research reads your site: What the logs reveal, Peec (David Konitzny), Jun 2026.