Fake Local News Sites Are Steering ChatGPT Answers
A network of 37 fake local news sites showed up in about half of ChatGPT's answers in NPR's test. Gemini never cited them. Here is how it worked.
👁 112 reads
The Kalamazoo Beacon looks like a small Michigan paper. Its tagline says it “covers Kalamazoo and lakeshore towns”. It has published 16 articles, and every one is about the same man: Sean McCann, a Democratic House candidate. When NPR asked ChatGPT about his district, the chatbot said McCann had the “most detailed” plan on cost of living - and cited the Beacon.
A network of 37 sites
The Beacon is one of 37 sites in a network uncovered by VisionHeight in a report published on October 7, 2026. All 37 domains were registered in mid-September in a single evening, 36 of them in under eight minutes. The structure repeats: for each of ten competitive districts there is one fake local news site and one opinion site, both promoting one Democrat. On top of that sit five aggregate sites and twelve state sites. The texts look AI-written; the phrase “is the stronger choice” appears more than 360 times. Who is behind the network is unknown.
What stands out is what the sites do not do. There are no fake quotes and no sensational lies. Instead there are candidate profiles of about 5,000 words and “voter guides” with 50-63 questions and answers, phrased the way people actually ask chatbots. The sites invite AI crawlers - their robots file names 17 of them, including OpenAI, Anthropic, Google and Perplexity - and mark their texts in code as news articles. Per NPR, each one also has an llms.txt file listing its articles in plain text. Most real outlets have no such file.
What NPR tested, and the limits
NPR’s Huo Jingnan published the results on October 30, 2026. The team asked 37 questions about the 10 districts, manually, on October 22-23, in ChatGPT, Gemini and Google Search, once each. That is a single run, not statistics, and NPR’s numbers should be read that way.
- ChatGPT cited at least one network article in 19 answers, roughly half. In 17 of those, the source was a fake local outlet.
- Gemini never cited a network site. In seven cases it declined to answer and pointed to outside sources.
- Google Search did not show these sites on the first results page. VisionHeight had found one there in early October, and Google said it acted on a policy violation.
After NPR alerted OpenAI, it repeated the McCann-district question five times. The Beacon appeared in every answer. OpenAI’s only response was that ChatGPT search cites a range of sources so people can make up their own minds.
The data void
The mechanism is not mysterious. Few people write about a single House district because local newsrooms keep shrinking. That leaves a data void. A chatbot answering a question about that district takes what looks local, current and well structured. It does not check who is behind a masthead. It checks whether the text fits the question - and these pages were built for exactly those questions.
GEO cuts both ways
Read the network’s playbook and it looks like a generative engine optimization checklist: let AI crawlers in, write in Q&A format, publish an llms.txt file, add schema markup. In this case, format beat reputation.
That has two uncomfortable implications for local media. First, a newsroom that blocks AI crawlers does not take itself out of chatbot answers; it leaves that space to others, including sites like the Beacon. Second, the method is cheap and portable. A few dozen domains registered in one evening plus machine-written text is within reach of almost anyone, in any market and any election.
What this means for local publishers
Local outlets have something the Beacon does not: real people, a real address and a track record that other outlets cite. The lesson from NPR’s test is that those signals only count if machines can read them. Bylines, an about page with names and contact details, and structured data that marks articles properly are no longer just good practice. They are what separates a real paper from a convincing shell in a chatbot’s eyes - assuming the chatbot looks.
How to check a chatbot’s source
- Click the citation. Chatbots can also invent citations, so make sure the page exists and says what the answer claims.
- Look for people. A masthead, names, a street address, a phone number.
- Count the articles. Sixteen pieces about one person is not a newsroom.
- Check the domain age. A WHOIS lookup shows the registration date. A “long-standing” local paper registered three weeks ago is a contradiction.
- Search the name in a normal search engine. Real local outlets cite each other.
- Ask a second model. In NPR’s test, Gemini and ChatGPT gave completely different sources.
The wider picture
VisionHeight also cites a NewsGuard audit: chatbots cited partisan sites in 48% of answers about the US midterms, and disclosed that partisanship in 1 answer out of 168. The Beacon is one example of a broader gap - chatbots treat a source as neutral unless something tells them otherwise.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources - checked on November 1, 2026
- NPR: Sites posing as local news boost candidates in ChatGPT answers (October 30, 2026)
- VisionHeight: Someone is manipulating your AI to swing your vote (October 7, 2026)
More on this from QUANTSLANT
news·Claude Code Agent in Pentagon System: WIRED ClaimsWIRED's anonymous sources say an Accenture contractor ran a personal Claude Code agent inside a Pentagon system. Accenture denies it; the DoD won't comment.