open (opened results) web-run/open

What it is

After searching, the model can open individual results: open|turn0search23. The reference points to a search hit (turn0search23 = 24th hit of the first search). ChatGPT then reads the page content; according to Segonzac as Markdown from a reading cache with 30-minute freshness, according to Konitzny in deep research mode in windows of ~5,700 characters followed by find as a text search. In our data open has so far appeared only in thinking runs, on 20 Aug in K5-01 thinking.

External evidence

• Olivier de Segonzac, SEL, 17 Aug 2026: citation rate opened vs merely retrieved.
• David Konitzny, Peec, 22 Jun 2026: Search → Open → Find, reading window, navigation density, first source 40.7 % of opens.
• Suganthan Mohanadasan, 24 Jun 2026: fetched / cited / mentioned as three stages.

For visibility

The funnel is: listed, opened, cited, visible. Each stage loses most of the previous one. Mohanadasan: 600 pages read, 30 cited. Segonzac: 74 % citation rate after opening vs 7 % without. To be cited you must be opened; to be opened you must convince in the excerpt (200 characters after the H1), not in the meta description. Konitzny also shows that the first hit receives 40.7 % of all opens and that pages with more than 60 links are read only one third as content.

External sources

segonzac-2026-08-17 · konitzny-2026-06-22 · mohanadasan-2026-06-24

Change history

Open

• In standard mode we saw not a single open across three days (opened_ref_count 0 in all 33 runs). Does only thinking/deep research open pages, or is opening invisible in the stream?
• What decides which hit gets opened — position, domain, excerpt?

Related

Query fan-out Retrieval pipelines (Labrador, scrape) sources_footnote