Skip to main content
ChatGPT’s Deep Research agent reads far more pages than it cites. Peec’s recorded session logs (full study, Jun 2026) show reading builds the agent’s understanding, while citing is a separate later decision it often declines — its own words: “the site doesn’t have specific content to cite for closures, so I’ll rely on general knowledge, as no citation is required here.” Consequences for measurement (Jun 2026 findings):
  • A page can shape the answer without appearing as a source. Tracking citations alone undercounts real influence.
  • Getting opened is necessary but not sufficient. To be cited you must be the specific, non-obvious source for a claim the agent decides needs backing.
  • Trust is graded before reading. The logs show consistent vocabulary: positive signals cluster around “official, reliable, trusted, credible, reputable”; user-generated and forum content is kept as background context, not primary evidence. The pattern mirrors E-E-A-T — be the official source in your category, or earn citations from it.
The retrieval-level analogue of this gap — retrieved but not cited — has its own metric and benchmarks: Retrieval vs citation. A case study of the gap at brand level: Kimi’s content strategy. Source: How ChatGPT Deep Research reads your site: What the logs reveal, Peec (David Konitzny), Jun 2026.