ChatGPT's Aura Check Can Block Sites Before Any Request Reaches Your Server

Ras Al Khaimah, United Arab Emirates, 30 October 2026.

ChatGPT's browsing tools can refuse to open a website even when its robots.txt, firewall and server all allow OpenAI in. An article published by Search Engine Journal on 29 October, written by an SEO consultant who advised an affected publisher, describes a central OpenAI safety check called Aura that decides whether ChatGPT's live search and Codex's Browser Use may load a site at all. OpenAI has not publicly documented how that check works. What it has documented is its crawler setup, and we pulled its published IP lists in this run to show where the line sits between a crawler you control and a block you cannot see.

What was reported

According to the article, Aura is an internal name that appears in OpenAI's systems and code for a site-level safety check. When a domain is flagged, tools that rely on the check decline to open it, and the refusal happens before any request leaves OpenAI. The author reports seeing status checks sent to a site_status endpoint on chatgpt.com, with a block result cached locally for 24 hours.

• The symptom: Browser Use returned the message "Browser use is not permitted on [domain]," and the agent was told not to look for workarounds.
• How long it lasted: the publisher in the case study first saw failures in May 2026, and all five URLs tested in a later check were still blocked.
• Controls: two rival news sites loaded in the same session, as did a page on another domain running the same technology. Setting the browser to always allow the site made no difference.
• Server logs: no ChatGPT-originated request reached the publisher's server at the times of the failed attempts.
• Support: OpenAI's front-line staff confirmed a safety block was active and escalated the case for review, but did not name Aura.

The cause of the block was not established. The article notes that some researchers suspect status requests that carry chat context could lead to site-wide flags, and labels that as unconfirmed. We treat it the same way: it is speculation, not a documented trigger. The article also cites OpenAI help pages describing a global blocklist for ChatGPT agent on security, safety and compliance grounds. Those help centre pages returned an access error to our test environment, so we have not verified their wording ourselves.

Crawlers you control versus a check you cannot see

OpenAI's crawler overview lists four user agents, each with its own published IP file. The page says each robots.txt setting works independently, that a site allowing both GPTBot and OAI-SearchBot may be served by a single crawl, and that ChatGPT search can take about 24 hours to reflect a robots.txt change. It also says ChatGPT-User acts on user requests, so robots.txt rules may not apply to it.

None of that touches the reported Aura check. A robots.txt rule or a firewall allowlist decides what happens when an OpenAI request reaches your server. A site-safety block decides whether the request is sent at all. That is why the case study found nothing in its logs: there was no request to log.

To make the visible side concrete, we downloaded all four of OpenAI's published IP files on 30 October 2026 and counted their contents.

User agent	Purpose (per OpenAI)	Respects robots.txt	IPv4 prefixes	IPv4 addresses	File timestamp
GPTBot	Crawls content that may be used for model training	Yes	18	2,704	22 Sep 2026
OAI-SearchBot	Surfaces sites in ChatGPT search	Yes	39	2,560	2 Jan 2026
ChatGPT-User	Fetches pages when a user asks	May not apply	235	69,280	27 Oct 2026
OAI-AdsBot	Checks landing pages submitted as ChatGPT ads	Not stated	2	256	12 May 2026

Three things stand out. First, ChatGPT-User has by far the largest footprint: 234 small /28 blocks plus a single /16 (9.129.0.0/16) that alone accounts for 65,536 of its 69,280 addresses. Its file was regenerated on 27 October, so allowlists copied even a week ago may already be stale. Second, six of GPTBot's 18 prefixes are identical to OAI-SearchBot prefixes, which fits OpenAI's note that one crawl can serve both purposes. A request from those ranges therefore cannot be pinned to one bot by IP alone; the user-agent string decides. Third, no ChatGPT-User prefix overlaps with either crawler, so a ChatGPT-User hit in your logs is a clean sign that live, user-triggered fetching works for your site.

For scale, Google publishes far more crawler ranges: our analysis of Google's crawler IP files counted 2,181 prefixes.

How to tell which problem you have

The practical question for an SEO team is whether ChatGPT cannot reach a site because of something on the site's side or something on OpenAI's side. The evidence points in different directions.

What you see	Likely meaning	Where to look
Requests from OpenAI ranges get 403, 429 or a challenge page	Your firewall, CDN or bot rules are blocking OpenAI	CDN and WAF logs, bot management settings
OAI-SearchBot fetches robots.txt but no pages	A robots.txt disallow is in effect	robots.txt, then wait about 24 hours after a change
ChatGPT says it cannot open the page and no request appears in logs at that moment	Consistent with an OpenAI-side block such as the reported Aura check	Timestamped tests plus log extracts for the same minute
Rival sites load in the same session but yours does not	Points away from a general ChatGPT outage	Control tests in one session

What SEO teams should do

1. Refresh your OpenAI allowlists from the source files. Pull chatgpt-user.json, searchbot.json and gptbot.json on a schedule rather than pasting ranges once. The ChatGPT-User file changed this week.
2. Test from inside ChatGPT, not only with a header spoofer. A curl request with an OpenAI user-agent string proves your server answers; it says nothing about whether OpenAI will send the request.
3. Match every failed test to your logs. Record the exact time and error text, then check CDN and origin logs for that minute. An empty log with a refusal message is the signature the case study describes.
4. Run controls. Test a few other domains, a subdomain and a deep URL in the same session, so you can show the problem is specific to your site.
5. Escalate with evidence and avoid workarounds. The article reports there is no public appeal route; it recommends opening support cases, keeping case numbers and asking whether a site-safety classification exists. It also warns that trying to bypass the block could breach OpenAI's usage policies.
6. Keep measurement separate. A blocked site loses visits that never show up as errors. Our earlier report on ChatGPT links losing UTM tags showed how easily ChatGPT referral data shifts for reasons outside a site's control.

Context

Most of the industry's attention on AI crawlers has gone to robots.txt. Our study of AI crawler blocking among 465 top sites looked at which bots publishers turn away, and Cloudflare has said it will score verified bots on robots.txt compliance. The reported Aura check runs the other way: the AI platform decides not to visit the publisher. Google offers Search Console to explain when it will not crawl or index a page. OpenAI has no comparable site-owner tool, and its public documentation covers its crawlers but not a site-safety classification.

Methodology

On 30 October 2026 we downloaded the four IP files linked from OpenAI's crawler overview (gptbot.json, searchbot.json, chatgpt-user.json and adsbot.json). For each we read the file's creationTime field, counted IPv4 and IPv6 prefixes, summed the addresses each prefix covers, and checked every pair of prefixes across files for overlap using Python's ipaddress module. Purposes and robots.txt behaviour come from OpenAI's crawler overview, read the same day. Details of the Aura check come from the Search Engine Journal article of 29 October; we could not test the check directly.

Conclusion

OpenAI's published crawlers are measurable and controllable: four user agents, four IP files and robots.txt tokens for the two automated crawlers. The reported Aura check is neither. It sits upstream of your server, leaves no trace in your logs and, on the evidence so far, has no documented appeal. Until OpenAI documents it, the most useful defence is evidence: current allowlists, timestamped tests and log extracts that show when a refusal came from OpenAI rather than from your own stack.

Questions and answers

What is OpenAI's Aura check?

According to a Search Engine Journal article published on 29 October 2026, Aura is an internal OpenAI name for a central safety check that decides whether ChatGPT's live search and Codex Browser Use may open a website. OpenAI has not publicly documented how it works.

Can robots.txt or a firewall allowlist override an Aura block?

Not according to the reported case. The block happens on OpenAI's side before any request reaches your server, so robots.txt rules and allowlists, which act on requests that arrive, do not affect it. The case study also found that a browser 'always allow' setting did not help.

How many IP ranges does OpenAI publish for its crawlers?

In files downloaded on 30 October 2026, OpenAI listed 18 IPv4 prefixes for GPTBot, 39 for OAI-SearchBot, 235 for ChatGPT-User and 2 for OAI-AdsBot. Six GPTBot prefixes were identical to OAI-SearchBot prefixes.

Does ChatGPT-User follow robots.txt?

OpenAI's crawler overview says ChatGPT-User acts on user requests, so robots.txt rules may not apply to it, and that OAI-SearchBot is the token to use for managing ChatGPT search opt-outs.

How can I tell if ChatGPT is blocking my site?

Test the page inside ChatGPT, record the exact time and error text, and check your CDN and server logs for that minute. A refusal with no matching request in the logs, while other sites load in the same session, is consistent with an OpenAI-side block.

###