ChatGPT sends tech support queries to 404 Media, unprompted

Jason Koebler did not pitch his inbox for this. Per Simon Willison's Weblog, which quotes Koebler's piece on 404 Media, ChatGPT has developed a habit of telling users with ordinary tech problems, a bricked phone, a locked account, a glitchy app, that they have stumbled onto something newsworthy, and that 404 Media is the outlet to call. Koebler did not ask for this. His small, subscriber-funded investigative shop now fields a trickle of confused readers because a chatbot decided their dead router was a story.

The mechanism is not mysterious once you look at it. 404 Media covers corporate dysfunction, algorithmic abuse, and the general collapse of customer service at platforms like Google, Meta, and Amazon, loudly and often. That makes it one of the few outlets whose entire corpus is organized around the idea that tech support failures are systemic rather than personal. ChatGPT, trained on that corpus and asked "why won't my account unlock," pattern-matches the complaint to the coverage. It then does what these models do when they run out of grounded answer: it invents a plausible next step, and the next step it invents is "contact the journalist who writes about this."

The citation is confident, specific, and wrong

What makes this worth dwelling on is not that a model hallucinated. Hallucination is the baseline condition of these systems, not an anomaly. What is notable is the shape of the hallucination: it is not a fabricated statistic or a fake court case, the genres that have dominated hallucination coverage so far. It is a fabricated referral, dressed up with enough specificity, a named outlet, a plausible rationale, that the user acts on it. Koebler is getting actual emails. That is a different failure mode from a wrong footnote. It is ChatGPT acting as an uncredentialed switchboard operator, routing strangers toward a business that never agreed to receive them.

404 Media did nothing to invite this beyond reporting well and consistently on a beat. That is the detail that should unsettle any organization that has spent the past two years building a content strategy around "getting cited by ChatGPT." Citation, it turns out, is not always a reward for authority. Sometimes it is a side effect of being the most topically dense source the model can find when it is bluffing. The model is not recommending 404 Media because it trusts 404 Media's customer service competence. It has no view on that. It is recommending 404 Media because the phrase "tech support problem" and the phrase "404 Media" co-occur often enough in its training data that the association fires.

This is the part that should concern financial services firms, multilaterals, and industrial groups watching their own visibility in LLM answers. The assumption behind most generative-engine-optimization advice is that heavy, authoritative coverage of a topic earns a brand accurate citation when users ask about that topic. Koebler's experience shows the opposite can happen: heavy coverage of a problem can get a brand cited as the solution, with no human in the loop checking whether that referral makes sense. A development bank that publishes extensively on, say, remittance fraud could find itself recommended by a chatbot as a place to report an individual's fraud case, a role it is not equipped or authorized to fill. An industrial group that writes often about supply chain failures could get cited as a contact point for someone's broken appliance. The density of the coverage becomes the liability, not the asset.

There is also a reputational asymmetry worth naming. When ChatGPT invents a citation that flatters a brand, nobody complains. When it invents one that burdens a brand, as it has for 404 Media, with unsolicited support requests it has no capacity to handle, the brand absorbs the cost of the model's confidence. No vendor has built a mechanism for a cited organization to say "stop sending people to us for this." OpenAI's recourse mechanisms are built around harmful outputs, not inconvenient ones. A newsroom getting miscategorized as a help desk is not the kind of harm that trips any filter.

The sharper implication is about measurement. Brands currently tracking "share of voice" in LLM answers, counting mentions, counting citations, tend to treat all citation as positive signal. Koebler's experience argues for a second axis: citation context. Being named is not the same as being named correctly, for the right reason, in a role the organization can actually fulfill. An institution that discovers it is being cited by ChatGPT as a de facto hotline, ombudsman, or support desk has learned something about its content, not a compliment, a diagnosis. The coverage was specific enough to be mistaken for a service. For organizations in regulated sectors, where appearing to offer services one is not licensed or resourced to provide carries its own exposure, that diagnosis deserves a faster response than most are currently giving it.