A block that leaves no trace. If ChatGPT and its agents decide your website is unsafe, they stop opening it, and not a single lost visit shows up in your logs. For leadership, the takeaway is simple: whether AI agents can reach your website is now a business metric, and it needs to be tracked the same way you track Google traffic.
What happened
An international publishing group with news and magazine content has been unable to get ChatGPT to load its pages since May 2026. Its servers are not the problem: the logs show the requests never even arrive. The cause is an internal OpenAI safety check, known as “Aura”, which classifies domains and decides whether its tools may open them.
Once a domain is flagged, ChatGPT’s live search and its agents’ browser (the one Codex uses to browse and act on the user’s behalf) refuse to go in. The block ignores user settings and robots.txt, is cached for 24 hours, and not even an administrator’s “always allow” rule lifts it. The agent itself is told not to look for workarounds.
The tests were clear: two direct competitors loaded without issue in the same session, and so did a third-party page built on the same technology but hosted on another domain. Weeks later, all five tested addresses (main site, subdomains and articles) were still blocked. OpenAI support confirmed in writing that a safety restriction was active and passed it to specialist teams to check whether it was “expected”. There is no status page, no appeal form and no timeline.
One figure explains why nobody noticed: over 30 days, OpenAI’s training crawlers made around 5,000 successful connections to the site, compared with 477,000 from Google and 564,000 from Anthropic. Traffic from OpenAI is small, and when it disappears entirely, nobody notices.
The bigger picture: a new gatekeeper that gives no notice
For twenty years, online visibility followed known rules. If Google did not index a page, Search Console told you. AI assistants bring a different gatekeeper: each platform decides, on its own and with little documentation, which websites its agents may consult. And it decides before the request ever leaves for your server, so no analytics tool records it.
This case involves a publisher, but the mechanism does not care about sector. An online store, a B2B catalogue or a corporate website could be flagged in exactly the same way, rightly or wrongly. The real danger is misdiagnosis: it is easy to assume the brand is missing from AI answers because of content or ranking, and invest in the wrong fix.
Why it matters to your business
AI agents already carry out tasks on request: comparing suppliers, finding spec sheets, building a shopping list. If your site cannot be opened, the agent does not stop to ask; it moves on to the next source. That means three things for the bottom line:
- Sales you never see lost. No failed visits, no errors, no abandoned carts. Your brand simply never makes the shortlist.
- Misallocated budget. If the problem is a block, no amount of investment in content or AI search optimisation will fix it.
- Dependence with no one to call. Unlike a traditional search engine, there is no clear channel to appeal today. Companies with a commercial relationship with the platform (advertising, enterprise agreements) will have a better chance of getting their case reviewed.
What to check this month
You do not need a technical project to find out whether this affects you. One morning from someone on the team is enough to answer these questions:
- Can ChatGPT open and summarise our home page, one category and three key product or service pages today? And can it open two competitors’ sites in the same session?
- If it fails, does our server log the request? If our logs show an error, the block is ours (firewall, anti-bot rules, geo-blocking). If there is nothing, the block is outside.
- Who keeps the evidence? Screenshots with date and time, the exact error message and the logs. That is what any support team asks for before escalating a case.
- Do we have a commercial contact at the platform who can push for a review?
And one rule worth stating clearly: do not try to get around the block. OpenAI’s usage policies forbid it, and the tool itself says so.
How we approach it at TRILOGI
For us, this news confirms something we already see in our projects: the AI assistant channel needs its own dashboard. Measuring sessions and conversions is no longer enough; you need to know, on a regular basis, whether each platform’s agents can open the URLs that sell the most, and cross-check that with the sales from those same pages.
Our approach: a short list of critical URLs (the ones that drive revenue), a scheduled access check from the main assistants, benchmarked against a couple of competitors as a control, and an alert when something changes. That indicator goes on the same dashboard leadership already reviews, next to traffic and sales by channel. If it drops one day, you know within hours rather than months, and you reach support with the evidence ready.
It is a small piece of our work in advanced analytics and AI dashboards: turning what happens in new channels into data you can make decisions with.
Every Tuesday, the 5 stories that matter: subscribe to the newsletter.
