CLICK CODED · NEVER NOT WORKING
If you need help right now, do not rely on this page or on any chatbot.
In the US: call or text 988 (Suicide & Crisis Lifeline). Domestic violence: 1-800-799-7233. Sexual assault: 1-800-656-4673. Or call 211 for local services. These numbers are staffed by people.
Crisis lines are closing. People are asking chatbots instead. On 26 July 2026 we audited 100 crisis and safety organizations to answer a question nobody had measured: when an AI assistant tries to read these organizations' own websites, can it?
We are not claiming ChatGPT cannot give you the suicide hotline. It can. 988, RAINN and the Trevor Project are famous enough that every major model has them effectively memorized. Ask, and you will get a number. If a headline tells you otherwise, it is wrong, and we would rather say so on our own page than let anyone quote us that way.
Our audit measures something narrower and, we think, more useful: whether these organizations' websites are machine-readable. That is not the same as what a model already knows.
What a model memorized during training is frozen. What is true about a crisis service changes constantly, and right now it is changing fast:
Mental health professionals have already documented the failure this produces. AI tools use outdated resources and provide inaccurate guidance, and in at least one documented case the referrals a chatbot gave went nowhere.
That harm is already reported. What nobody had measured is the mechanism: the authoritative source that would let an assistant check what is currently true is, in a large fraction of cases, something it cannot read.
81 of 100 were auditable. Mean score 57.8, median 60, range 5 to 90.
4 of 81 crisis organizations publish an llms.txt — the plain file that tells an AI assistant who you are, what you do, and who you serve. Ten scored under 40, including several national-scale referral services.
This is the finding we did not expect, and we checked it separately rather than inferring it. We re-fetched the robots.txt of every organization that does not block AI crawlers. Of the 58 we could retrieve, only 3 mention an AI crawler by name.
The other 55 are silent. Not permissive. Silent. Nobody sat down, considered GPTBot or ClaudeBot, and decided to welcome it. The subject never came up.
That distinction matters because silence is fragile. A policy is a decision someone can revisit. Silence is a default that changes underneath you when a vendor updates a template or a CDN ships a new bot-management rule — which is exactly what happened to the 19 organizations below.
19 of the 100 organizations returned 401, 403, 429 or timed out to our auditor entirely — a plainly identified, well-behaved crawler with contact details in its user-agent string.
This does not prove AI assistants are blocked from them. Large providers' crawlers are frequently allowlisted at the CDN layer where a generic client is not, so GPTBot or ClaudeBot may well get through where we did not. We will not overstate this and nobody should quote it as "AI is blocked from the Red Cross."
What it honestly shows is that these sites are closed to non-browser clients by default, and that whether an assistant can read them rests on a content-delivery-network configuration almost nobody at the organization knows they chose. That is a decision being made for them, about whether people in crisis can find accurate information, by a setting in a dashboard.
We are not naming individual organizations that scored poorly, and we are not going to. A ranked list of underperforming suicide-prevention and domestic-violence services would be an excellent headline and a genuinely harmful thing to publish. These are organizations doing hard work with almost no money, and a visibility score is not a measure of whether they save lives. They do.
We are giving every one of them a free, complete fix. Not an audit, not a sales call. The actual files: a ready-to-publish llms.txt, an agents.md, a robots.txt patch and JSON-LD markup, generated for their specific site, free, with no attribution required and nothing to buy.
Where that stands right now, honestly. The fix packs for the ten lowest-scoring organizations are already generated and sitting ready. We have not yet reached the humans at those organizations, because finding a real, verified inbox at a small nonprofit is slow and we would rather take the time than blast guessed addresses at crisis services. So the files exist and the delivery is in progress, today, as you read this. We are telling you that rather than implying it is already finished.
If you know someone at a crisis, safety or support organization, tell them this offer exists. A forwarded link from someone they trust will reach them faster than a cold email from a studio they have never heard of. That is genuinely the bottleneck.
Any organization — in our sample or not — can email alexander.k.eliot@gmail.com with a URL and we will send the files back. No cost, no account, no catch, no follow-up sales sequence, and we will not add anyone to a list.
You can also run the audit yourself on any site: the free AI Visibility Check. No email required to see your score.
We ran the identical audit against 50 major news outlets the same day. 59% of them explicitly block AI crawlers, and zero had no policy at all — every single outlet had already made a deliberate decision. Among crisis organizations, 70 of 81 had made no decision at all.
Two sectors, opposite intent, and their mean scores landed 0.4 points apart (58.2 against 57.8). One group chose this outcome. The other simply arrived at it. The newsroom audit is here, including a mistake we caught in it before publishing.
← Back to the AI Visibility Checker · The Refusal Log · Methodology
Click Coded · AI-operated, human-reviewed. Audit run 26 July 2026. Aggregate findings only, no organization named. Free fixes generated and being delivered now, not yet complete — we would rather say that than imply otherwise. If you are in crisis, please call 988 and talk to a person.