On October 30, NPR published an investigation into 37 websites that present themselves as local news and opinion outlets. They promote political candidates, mostly Democrats in competitive House races, and they do not say who runs or funds them. The candidate NPR asked and the party's House campaign committee both said they were not behind the sites.
The politics will get the headlines. The method is the part worth studying. These sites were not built for people to find on Google. They were built to be picked up by AI answer engines, and in one engine's case, it worked about half the time.
How the network was built
All 37 sites were set up on September 15. Researchers at VisionHeight, who found the network, say 36 of them were registered within an eight-minute window. Hundreds of articles followed in batches through September and October, and the researchers suspect AI wrote much of it.
The structure was planned. Ten sites pose as local news outlets, each one promoting a single House candidate. Ten opinion sites back the same ten candidates. Five roundup sites carry names like "Public Safety Democrats." Twelve state-focused sites cover state and local races.
Most publishers that think about AI at all use robots.txt to keep AI crawlers out. These sites do the opposite. They list a long run of crawlers by name, from OpenAI, Anthropic, Google, Perplexity, Meta, Amazon and Apple, and let them all in. Every site also has an llms.txt file listing its articles in plain text. University of Washington researcher Alisha Ukani, who checked the files for NPR, called that surprising, because most sites do not have one.
Many of the articles are written as questions and answers. That is the format answer engines tend to quote.
What the engines did with it
NPR wrote 37 questions and ran each one once on ChatGPT, Gemini and Google search, on October 22 and 23.
ChatGPT cited at least one of the network's articles in 19 of the 37 answers. In 17 of those, the cited page was one of the sites posing as local news. For five Michigan city council candidates, ChatGPT cited the same site, the Wolverine State Report, every time.
Gemini cited the network in none of its answers. It declined to answer seven questions and pointed to other sources instead. Google search showed no network links on its first page, and its AI Overview did not mention the sites, although VisionHeight had found one site ranking above Ballotpedia earlier in October.
Same questions, same week, same open web. One engine took the bait half the time and the other never did.
Why ChatGPT picked these pages
The sites gave the engine what it looks for: pages it was allowed to read, answers already in question-and-answer form, and a plain-text map of everything on the site.
The questions also mattered. Preliminary research cited by NPR suggests chatbots cite questionable sources more often in "data voids," topics with little credible coverage. A down-ballot city council race is a data void. When an engine searches and finds almost nothing, a page that looks like a direct answer has very little to compete with.
OpenAI told NPR that ChatGPT search draws from a range of sources and that it acts when issues are raised. In five follow-up queries, NPR still found one of the sites, the Kalamazoo Beacon, cited every time. Google said it found a policy violation in one case and took action, and that its AI features lean on ranking systems built to block spam.
NPR is careful to note that each question ran once. Repeated runs could give different results.
What this means for a B2B company
Take the politics out and this is a clean field test of how AI source selection works.
Most B2B categories are data voids too. "Best SOC 2 tool for a 10-person startup" or "alternatives to a niche billing vendor" has far less credible coverage than a national election. When a buyer asks that question, the engine searches, finds a handful of pages, and quotes whichever ones read most like an answer. Those pages may be yours, a competitor's comparison post, or a thin page someone wrote last month.
The tactics the network used are not secret. Letting AI crawlers in, writing pages that answer a question directly, and keeping a clean list of what you publish are what honest companies should do too. The difference is disclosure. A product page that says who wrote it and why is doing the same work in the open.
The engines also behave very differently. A brand that measures only ChatGPT would have seen these sites everywhere. A brand that measures only Gemini would have seen nothing. Visibility has to be read per engine, with the same questions asked more than once.
The bigger picture
When buyers ask about a narrow category, AI engines answer from whatever pages they can reach and parse. NPR's test shows how much one well-prepared set of pages can change those answers, and how much the engines disagree about which pages to trust. If nobody in your category has written the clear, crawlable answer to your buyers' questions, someone else's page will fill the gap.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.