Key points
- An SEJ contributor reports that OpenAI runs a site-safety check, called Aura, that can stop ChatGPT live search and Codex Browser Use from opening a site.
- The block overrides robots.txt and user settings, is cached locally for 24 hours, and has no public appeal or review process, according to the report.
- One international media publisher has reportedly been blocked since May 2026, while competitor sites loaded in the same sessions and server logs showed no ChatGPT requests.
- OpenAI support reportedly confirmed a safety block was active and escalated the case to check whether it was intended.
Why it matters for marketers: A hidden platform-side block can erase your brand from ChatGPT agent answers while analytics show nothing, so GEO and AEO spending may be wasted until you test agent access.
A Search Engine Journal contributor describes a site-safety check inside OpenAI's systems, referred to as Aura, that decides whether agent tools may open a given website. If a domain is flagged, ChatGPT live search and Codex Browser Use reportedly refuse to load it, regardless of robots.txt rules or local permission settings. OpenAI's help pages mention a global blocklist for security, safety and compliance, but the author says the mechanics and the criteria are not public.
The evidence comes from a publisher the author advises. Since May 2026, ChatGPT reportedly failed to load its pages, and server logs showed no incoming requests, which points to a block before traffic leaves OpenAI. Competitor news sites opened in the same sessions, and a third-party page running the brand's technology on another domain also loaded. OpenAI support reportedly acknowledged a safety block and escalated the case to check whether it was expected.
For marketers, the risk is that this looks like a visibility problem when it is an access problem. The author argues that a blocked site is simply skipped by agents, and ChatGPT may lean on secondary sources or outdated coverage instead. Because the failed visits never reach your servers, analytics will not show the loss. The author also notes that advertisers could pay for placement while agents cannot open their links.
The practical step is to add agent access checks to AI search audits. The article suggests testing the same URLs across ChatGPT fetch and Codex, running competitor controls in the same session, comparing server logs with crawler analytics, and ruling out your own firewall and bot rules first. If the block looks upstream, build an evidence pack, open linked support cases, and avoid workarounds. This is one contributor's account of a single case, and OpenAI has not publicly detailed the system.
Sources
This brief is our own summary. Read the full reporting at the source.
