Summary
- OpenAI’s ChatGPT cited articles from a 37-site undisclosed promotional network in 19 of 37 NPR test queries because citation ranking evaluates content-level signals that batch-registered properties can reproduce at scale.
- Registry data gathered by VisionHeight and Andreas Sjöstedt shows all 37 sites were set up on September 15, all but one within an 8-minute period, suggesting one templated operation rather than independent outlets.
- The network’s permissive AI-crawler settings, llms.txt files and question-and-answer article formatting appear engineered for chatbot pickup.
- A “data void” in down-ballot races amplifies the network’s reach, a mechanism that preliminary research cited by NPR links to chatbots citing questionable sources more often where credible reporting is scarce.
- OpenAI’s post-notification remediation showed no immediate retrieval change: the Kalamazoo Beacon “was still cited as a source every time” across five re-test queries.
An NPR test found that ChatGPT, OpenAI’s chatbot, cited at least one article from a network of 37 websites that pose as local news in 19 of 37 queries — about half — while Google’s Gemini cited none. The network’s most visible placement came in Michigan’s 4th Congressional District, where ChatGPT put Democrat Sean McCann’s “Most detailed published affordability plan” at the top of its answer and cited the Kalamazoo Beacon, a site that has published 16 stories, all about McCann. The network’s success traces to a systemic gap: its synchronous registration, uniform file structure and retrieval-oriented formatting reproduce the content-level signals that citation ranking systems use to infer credibility, and nothing in the pipeline distinguishes 37 synchronously registered properties from 37 independent newsrooms. Registry data gathered by VisionHeight and Andreas Sjöstedt shows all 37 sites were set up on September 15, all but one within an 8-minute period. Who operates the network remains unidentified: NPR was unable to find or reach the site operators, candidates and committees it contacted denied involvement or did not respond, and the effect reached down ballot, where ChatGPT cited the Wolverine State Report every time NPR asked about the positions of five Michigan city council candidates.
The test and the network
Less than two weeks before Election Day, NPR asked ChatGPT which House candidates in Michigan’s 4th Congressional District care about affordability. The chatbot put McCann’s “Most detailed published affordability plan” at the top of its answer and cited the Kalamazoo Beacon; the race is one of the competitive U.S. House contests that could determine control of the chamber. To gauge the network’s reach, NPR drafted 37 questions and used them on ChatGPT and Gemini, the two most popular AI chatbots, as well as on Google search, which sometimes provides AI-generated answers. For each of the 10 congressional races, NPR asked general questions — such as which House candidates care about affordability or public safety — along with questions specific to individual candidates. The queries ran manually on October 22 and 23, with a VPN changing the IP address to the corresponding state or a nearby state and a separate private browsing session for each tool and prompt. In 19 responses, about half, ChatGPT cited at least one article from the network, often alongside other sources such as Ballotpedia or the candidates’ own campaign websites. In 17 of those 19, the chatbot cited one of the outlets posing as local news; in the other two, it cited the network’s state-focused websites. NPR noted that chatbots can give significantly different responses to the same question across submissions, and submitted each query only once, so parallel or future experiments may yield different results.
The network’s reach extended down ballot: NPR identified five Michigan city council candidates covered by a site called the Wolverine State Report, and ChatGPT cited that website every time NPR asked about the positions of each of the candidates. Google’s products behaved very differently: Gemini cited no links from the group in any response, declined to respond at all in seven instances, and Google search did not surface any of the network’s links in its first page results, nor did its AI Overview mention them. Google spokesperson Davis Thompson said the company detected a policy violation in one instance — VisionHeight’s report showed that in early October at least one site had appeared on Google’s first page, ranking higher than Ballotpedia — and took “swift action.” “Our generative AI features are built on Search ranking systems, which are designed to show high quality content and block spam and abuse,” Thompson said. VisionHeight’s CTO, Nitsan Daniel, said the firm did not systematically measure how the sites showed up for different candidates.
The Kalamazoo Beacon labels its articles as “news articles” in the website’s source code and describes itself as “Reporting on Kalamazoo, southwest Michigan, and the issues shaping lakeshore communities.” NPR’s review shows it has published 16 stories, all about McCann, his policy positions and his record as a state legislator. The Beacon is one of 10 “local news” websites that promote one Democrat apiece in competitive U.S. House races — part of a larger group of 37 websites identified by VisionHeight that promote many more candidates, mostly Democrats, for federal, state and local offices, as well as some nonpartisan posts like judges. The 37 websites present candidates in overlapping groups, NPR’s review found. Each of the 10 Democrats covered by a local-news site is also the sole subject of a site that offers opinion articles about why voters should support them. The same candidates appear in roundup-style articles on five more websites with names such as “Public Safety Democrats” and “Democratic Women in the House,” and another 12 state-focused websites covered the House candidates along with other state and local races. The state and national websites also boosted an 11th House candidate who did not have her own dedicated site purporting to be local news. While most of the sites are openly partisan, the articles link out to sources for the claims they make — congressional voting records, mainstream news articles and partisan sources such as the pro-Democrat Courier network. Opposing candidates mostly made infrequent and brief appearances; articles that engaged with them substantively focused on possible political liabilities such as support for data centers, not stopping the Iran war, rolling back Medicaid and supporting tariffs.
Evidence of a single operator
Multiple indicators suggest all 37 websites share a common operator. All were set up on September 15 — all but one within an 8-minute period — according to registry and digital certificate data gathered by VisionHeight and Andreas Sjöstedt, a co-founder of the hybrid threat forensics firm Erebus who was not involved in VisionHeight’s research. Hundreds of articles appeared on the sites in batches in September and October; VisionHeight’s report said the volume suggests the operator may have used AI to fill out the material. Earlier in October, the 10 local-news outlets shared the same description — “Straight local news - reports on the race and (once a content-sourcing workflow exists) community storylines, in neutral third-person news style.” Since then, the descriptions have become more tailored to the areas the outlets purport to cover. One-day registration of all 37 sites, all but one within an 8-minute period, plus that shared placeholder description, indicates templated, low-marginal-cost production rather than per-outlet reporting.
The sites also appear built for chatbot pickup. All of them explicitly allow a long list of AI crawlers made by major AI companies to scrape their content, and many articles are formatted as questions and answers, which chatbots tend to prefer to cite; many news organizations block such crawlers because they take content without bringing in revenue. Every site in the network contains a file called llms.txt that lists articles in plain text — a pattern Alisha Ukani, a postdoctoral researcher at the University of Washington’s Center for an Informed Public who noticed the files while reviewing the sites at NPR’s request, said is unusual, since most websites don’t have such files. In an email, she noted that the websites that do have llms.txt files tend to be “biased towards technology and business-related websites.” The architecture of the operation is visible; its operator is not.
Root cause: the absent provenance signal
The immediate phenomenon is a chatbot citing undisclosed political advertising formatted as local news. The first-order cause is that the sites were retrievable and semantically on-point while credible alternatives were absent or blocked. The structural conditions are not of this operator’s making: preliminary research cited in the article found chatbots tend to cite questionable sources more often “when there are very few sources of credible reporting” — a phenomenon called a “data void” that is characteristic of down-ballot races, where there is often less news coverage — and legitimate outlets block crawlers for revenue reasons.
Citation systems evaluate content-level signals — question-and-answer formatting, links to congressional voting records and mainstream articles, neutral third-person style — all of which the network reproduced. Domain-level credibility is inferred from those signals, and they are cheap to synthesize at scale across 37 properties. Nothing in the pipeline distinguishes 37 synchronously registered properties from 37 independent newsrooms.
At the systemic level, the disclosure norm that governs the offline equivalent has no machine-readable enforcement point. Renée DiResta, an associate research professor at Georgetown University who studies online influence campaigns, said it makes sense that political campaigns and their allies would try to optimize their message for pickup by AI chatbots, but “pretending to operate as news outlets crosses a line into deception.” “Disclosing funding … would be obviously the ethical thing to do here, or at least indicating that this is a linked effort,” she said. The volume of websites and articles may also help the network’s operator boost preferred narratives, she said. Detection currently runs on external flagging — OpenAI’s spokesperson said the company takes action “when issues are brought to our attention.”
The data void and the crawler-access asymmetry amplify the probability of this outcome but do not originate it; question-and-answer formatting and publication volume amplify it further. What permits recurrence is the absent provenance and disclosure signal in citation ranking. A fix aimed only at the 37 identified sites would not prevent a 38th.
Who benefits
The episode’s benefits are concentrated and its costs are diffuse. The network’s most direct beneficiaries are the candidates it promotes, through retrieval placement: the top-of-answer “Most detailed published affordability plan” distinction for McCann, sourced to a site presenting itself as local news, and the repeated citation of the Wolverine State Report for five city council candidates. Those placements carry no sponsorship disclosure that would let an audience discount them.
The same absence of disclosure creates association risk for the promoted candidates and committees, consistent with the denials and non-responses NPR recorded. NPR was unable to find or reach the site operators. JoAnna Mendoza of Arizona, one of the Democrats the network promoted, said she wasn’t behind the websites and doesn’t know who is. The Democratic Congressional Campaign Committee denied involvement in a phone call with NPR. Other candidates the websites promoted, including McCann, and Democratic super PACs that support them — American Bridge 21st Century, Tomorrow Together and House Majority PAC — did not respond to NPR’s requests for comment. Denials and silence are unverified statements by interested parties, and the operator’s silence is not more probative than the committees’.
On NPR’s testing, Google’s search and assistant products did not surface the network, and Thompson said the company acted on the one documented violation. OpenAI’s response on the record: spokesperson Nate Evans said, “ChatGPT’s search tools cite from a range of sources so people can explore and make their own judgements,” adding that “When issues are brought to our attention, we take action to improve our model responses.” The post-flag record shows no immediate retrieval change: it’s unclear what OpenAI has changed since NPR flagged the sites, and the Kalamazoo Beacon “was still cited as a source every time” after NPR received OpenAI’s response and ran the query five times. In some instances, McCann no longer received the “most detailed affordability plan” distinction; in others, he still does. Report-driven remediation had not yet changed retrieval behavior at the time of testing.
A third benefit pathway runs to the organizations that surfaced the episode: VisionHeight published the identification, Sjöstedt contributed registry and certificate analysis, and researchers at Georgetown and the University of Washington supplied interpretation.
The distributional costs fall on voters who read a cited answer without evaluating the underlying sources; on competing candidates, whose appearances NPR found infrequent and brief; on down-ballot communities, where the credible reporting pool is thinnest; and on news organizations whose crawler-blocking narrows the citation pool while the network’s permissive settings widen its own access.
Competing explanations
The record excludes the null reading: 37 sites registered on the same day, all but one within an 8-minute window, sharing description text and file structures, is not plausibly organic. Common-date registration and the shared placeholder description are highly diagnostic of a single coordinated operator and weigh against independent, organically grown local outlets. The technical scaffolding — crawler allowances, llms.txt files and question-and-answer formatting — is diagnostic of a build aimed at chatbot pickup specifically, separating the operation from ordinary partisan blogging.
The coordinated-undisclosed-promotion reading is the strongest but remains supported rather than sealed, because the operator is unidentified. A narrower reading is also consistent with the record: most of the sites are openly partisan and only the 10 local-news outlets pose as independent journalism, so the operation may be a hybrid — an openly partisan booster effort with a disguised news layer. DiResta’s characterization is directed at that disguised layer.
A party-committee or super-PAC operation explains favorable coverage of Democrats but fits less well with anonymous operation, nonpartisan judicial races, and a candidate who denies involvement. A candidate- or campaign-aligned vendor explains the single-subject local sites but not the multi-candidate roundup sites. A foreign or otherwise state-linked operation is an explanation the article’s evidence does not directly support: it reports no attribution evidence, the unfinished placeholder description suggests a work in progress rather than a mature operation, and the promoted candidate set aligns with domestic partisan objectives; VisionHeight’s volume statement is consistent with resourcing.
The underexplored hypothesis is a commercial or promotional intermediary selling a packaged, AI-visible presence across many races at once. Batch registration, uniform templates, and a portfolio mixing partisan and nonpartisan contests are the pattern a vendor’s product would leave, and the AI-optimized production pattern is also characteristic of commercial content operations generally; the article reports no evidence of the operator’s identity or business model. The falsification path for the intermediary hypothesis is concrete: shared hosting or a common provider across the 37 domains, matching WHOIS privacy-service patterns, common payment or registration rails, and byte-identical template fingerprints would each confirm it, and none present would weaken it. Two fingerprints are already in the record — the shared placeholder description and the uniform llms.txt file — while the article supplies no hosting, payment or contractual evidence, so the branch remains open rather than established.
Sensitivity runs in both directions. An admission or documented linkage by a committee would move coordinated party operation to the front; disclosure of a vendor contract would move the intermediary hypothesis ahead of both; if the DCCC’s denial were withdrawn, the case would rest on documentation rather than inference — the current position of every hypothesis, since none is established. The comparison is also bounded: in the other 18 of the 37 responses, the article does not record whether credible local coverage was present, whether the model declined, or which sources it used, so a within-sample contrast against the data-void mechanism cannot be drawn from the reported data alone; the mechanism is carried by the preliminary research the article cites rather than by variance within NPR’s own sample. The readings are not mutually exclusive: a coordinated political operation can also be a commercial one, and the most fitting combination is the production mechanism with the data-void mechanism — coordinated content is most effective where the pool of credible reporting is thinnest.
How the episode is being framed
The episode supports three readings, each assigning a different remedy. As a citation-quality problem, the remedy sits with AI companies and their ranking systems. As an advertising-disclosure problem, it sits with the operators and the disclosure DiResta identified as absent. As a data-void problem, it sits with the availability of credible local reporting. The record supports all three at once.
The account does not establish that the retrieval-optimized format itself — machine-readable listings, question-and-answer formatting, source links, permissive crawling — is the objectionable element. The disclosure standard concerns sponsorship, not format; whether a sponsorship-labeled version of the same architecture would function differently is not addressed in the record. The element identified as missing is the disclosure.
The underlying practice of making a candidate’s record and positions retrievable and citable is one that campaigns perform openly through websites and issue pages, which ChatGPT also cited in NPR’s testing. What the network added is the distributional overlay — coordinated volume, machine-readable exposure, and a news form borrowing the credibility of independent journalism without disclosing sponsorship that would let an audience discount it.
The same technical footprint contains no partisan signature: batch registration, permissive crawler settings, machine-readable listings and question-and-answer formatting are available to any coordinated content operation, and the researchers quoted in the article describe the phenomenon in general terms. The same analytical standard would apply to a network serving any other candidates; the specific factual record concerns a network promoting Democrats. DiResta also described a “competition for curation,” with opposition research entities working to make sure bad information about candidates is found — the adversarial counterpart to the network’s promotional design.
Patrick Warren, a professor at Clemson University’s Media Forensics Hub, put the episode in lineage: thinly veiled advertising that purports to be a news source has a long history — from ad flyers made to look like newspapers to ads mixed in with legitimate news in traditional online search. Reading AI chatbot answers without evaluating the underlying sources is the latest version of that experience, he said. “There’s nothing new under the sun.” What is new is the mechanism.
What remains open
NPR’s review was based on downloads of every website homepage and article page twice, on October 19 and October 24, records also used to verify VisionHeight’s observations. NPR manually reviewed all the “local news” and “local opinion” sites, used AI to analyze articles on the sites focused on state and national races, and conducted keyword searches and manual reviews to examine how opposing candidates were portrayed.
Three questions remain open on the public record: who operates the network; what OpenAI has changed since flagging; and how the competitive-curation dynamic develops. On the last, DiResta said the “competition for curation” “is going to have interesting ramifications as the frontier companies themselves try to figure out how to get better at not falling for it.”
Independent external verification of the article-supplied figures — the registration-window data, the 16-story count, the 17-of-19 citation count and the post-notification re-test finding — was not available at the time of writing; the figures stand as reported in NPR’s account.
Analytical techniques used in this piece
This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.
- Cui Bono — Who Benefits
- Asks who gains and who pays from a state of affairs, decision, or claim.
- Differential Diagnosis
- Lists the candidate explanations for a symptom and rules them out one by one.
- Root-Cause Analysis
- Traces a symptom back along its causal chain to the conditions that actually generated it.
- Signaling
- Costly, hard-to-fake actions that credibly communicate intent or quality.