Knowledge baseSourcesLily Ray

What we can learn from evolving ChatGPT search

Author Lily Ray Date 2026-08-17 Model state Own observations plus Promptwatch data, August 2026 Open source →

Key findings

Framing of the site: switch for SEOs: which domain types are preferred, the risk of the model confusing domains.

Checked against our data

The domain types match our lists, with one difference: Reddit does not appear for German prompts. The lists are short and, per question, largely stable across two days.

ClaimStatusEvidence
site: share ~23%, retrieved sources rose from ~12 to 24outdatedIn our data 14-22 URLs per searching question (search.url_count), domain restriction in 58% of searching questions.
Narrowing to .gov, review platforms, manufacturers, Redditconfirmedcdc.gov, heart.org, nhs.uk (health), apple.com/samsung.com/dyson.de/gaggia.com (manufacturers), idealo.de (prices), leading media (news). Reddit never.

Related

promptwatch-2026-08-10