Knowledge baseSourcesMetehan Yesilyurt

Reddit lost most of its visibility in ChatGPT, and the site parameter filled the gap

Author Metehan Yesilyurt Date 2026-08-21 Model state GPT-5/5.6 era, own query corpus (14.9M queries in the peak week), as of 21 Aug 2026 Open source →

Key findings

Yesilyurt dates the Reddit collapse in ChatGPT citations to two steps (8 and 14 Aug 2026) and shows, in parallel, the explosion of the site: parameter from practically zero to 26.2% of all queries - aimed at official sources, almost never at Reddit. His reading: not a Reddit problem but a shift in retrieval strategy, in which targeted domain restriction displaces broad community hits.

Why this source matters for the project

The two-step dating calibrates the era our baseline starts in: we have been measuring the post-Reddit order from day one. And it connects two previously separate register entries - Reddit decline and domain-filter rise - into one movement.

Checked against our data

Both change points predate our first run - exactly the gap the tracker closes from now on. From 20 Aug our findings match his end state: no Reddit in the hits, domain restriction as the norm (58% of searching standard questions with a domain filter) on official sources. He extends Mohanadasan's Reddit figure (3.83% to under 1%) with the more precise dating and the counter-movement.

ClaimStatusEvidence
Reddit citations fell in two steps: around 8 Aug from 4.4-4.6% to ~3%, on 14 Aug a cliff to 0.55%, at 0.23% by 21 Aug (-95%)unverifiableBoth steps predate our measurement start (20 Aug). Consistent: Reddit appears in none of our five measurement days' hits - German prompt set, no Reddit prompt.
The site: share rose from under 0.11% per week to 10.2% (week of 3 Aug) and 26.2% (week of 10 Aug, 3.9M of 14.9M queries)unverifiableHis corpus, his scale. With us the mechanism exists twice: site: in the query text ([metric/fanout-site-share], sporadic) and the pipe domain parameter ([metric/fanout-domain-filter-share], 8 of 18 searching standard questions on 22 Aug).
site: queries target official sources, under 1% RedditconfirmedOur domain-filter targets are official or established sources throughout (openai.com, reuters.com, tagesschau.de, britannica.com, dw.com) - never Reddit.
The cause is presumably a cache/index change at OpenAIunverifiableServer-side; not observable in the client capture.

Related

promptwatch-2026-08-10 mohanadasan-2026-08-21 yesilyurt-rrf