Russia runs a Swiss news site. Almost nobody reads it, and that is the point.

On the morning after Switzerland voted down the Neutrality Initiative, I opened a Swiss news site that almost nobody in Switzerland has heard of. By noon it had published around twenty articles, ten of them about Sunday’s votes. “Switzerland has taken off its neutrality mask,” said one headline. “The Russian ambassador said Switzerland has no neutrality,” said another.

The site is switzerland.news-pravda.com. It is not Swiss, it has almost no readers, and that is not a failure of the operation. It is the design.

What switzerland.news-pravda.com published the morning after the vote

The vote itself was not close. 70.15 percent said no, and not a single canton said yes.

The site treated that result as a story about Swiss betrayal. Between 07:00 and noon on 28 September it ran pieces calling the result proof that Switzerland is no longer neutral, a quote from Russia’s ambassador in Vienna, and several rewrites of the same agency report. The newest article that morning carried the number 20637 in its URL.

The Swiss site on 28 September 2026: five of these headlines are about the vote. Note “Außenminister” with ß. Screenshot by the author.

It is one of more than 80 subdomains that branch off news-pravda.com, one per country or theme: austria., germany., nato., trump., and so on. France’s agency against foreign interference, VIGINUM, first documented the network in February 2024 under the name Portal Kombat, counted 193 sites at the time, and tied its setup to TigerWeb, a web company based in occupied Crimea.

We are not linking to it. A link is a vote for a page in the eyes of a search engine, and this network exists to collect those.

How Russia publishes thousands of articles a day

The American Sunlight Project estimated the network’s output at about 20,000 articles every 48 hours, or roughly 3.6 million in 2024. Nobody writes that much. The articles are assembled, and the pipeline is visible from the outside if you read a few of them:

The mirroring step is the one I have not seen described elsewhere. Every article we opened linked to a copy of itself on telegra.ph, a publishing tool run by Telegram. That doubles the number of indexable pages for no extra effort and puts a second domain behind every claim.

The source line is printed under the article: RT DE, a broadcaster banned in the EU since 2022.

How to recognise a Pravda article

Machine translation leaves fingerprints, and this site is full of them.

One article about the ambassador ends with the single word “Unterschreibe!”, which is a literal translation of “Subscribe!” and means “Sign it!” in German. Its source line reads Quelle: MAX "lomovka", a channel on Russia’s state-backed messenger. The text uses “ß” throughout, which Swiss German dropped decades ago, so a site calling itself Swiss news is written in a spelling no Swiss newsroom uses. Another piece says Switzerland voted to “die Seite zu besetzen”, a word-for-word rendering of a Russian phrase that no German speaker would write.

“Unterschreibe!” is a machine translation of “Subscribe!”. The source is a channel on MAX, the Russian messenger.

None of these tells is proof on its own. Together, and combined with the domain, they make a site easy to place.

Why a propaganda site does not need readers

NewsGuard pulled traffic estimates for two of the English sites: roughly 955 monthly visitors for one and about 1,000 for another. RT.com, by comparison, had around 14.4 million. By human standards the network is a failure.

It is not aiming at humans. Three kinds of machine read these pages, and all three matter more than a visitor would:

1. Search indexes. Millions of pages across dozens of languages mean that on a narrow or breaking topic, a Pravda page is often one of very few pages that exist.
2. Wikipedia. DFRLab and CheckFirst found 1,907 links to 162 Pravda sites across Wikipedia pages in 44 languages. Every one of those links tells a reader, and a crawler, that the page is a citable source.
3. Training data. In April 2026 DFRLab reported that Common Crawl, the public web archive that most large AI models draw on for training, held about 40,000 English-language Pravda articles by November 2025. A year earlier it held 37.

The researchers call this LLM grooming: flooding the open web so that the next generation of models learns from it. That is what the telegra.ph copies and the thousands of daily articles are for.

Does Russian propaganda actually change what ChatGPT says?

This is where the story gets less tidy than most coverage makes it, and where it is worth being precise.

In March 2025 NewsGuard tested ten leading chatbots with prompts built around false claims the network had pushed. Collectively they repeated the claims 33 percent of the time. That number went around the world.

In October 2025 a study in the Harvard Kennedy School Misinformation Review collected 416 answers from four chatbots, some of them queried from Switzerland, and found 5 percent supporting Kremlin disinformation. About 8 percent cited a Pravda site at all. The authors argue the cause is not grooming but data voids: the chatbots cited Pravda pages on questions so narrow that little else had been written about them. Critics of the NewsGuard study also point out that its prompts were designed to provoke the claims.

Both findings can be true at once, and together they point the same way. A chatbot is most likely to repeat propaganda exactly where credible reporting is missing. A Swiss vote, covered in German, French and Italian by a small number of newsrooms, is that kind of topic. On 28 September the Swiss Pravda site had nine fresh pages about the Neutrality Initiative before lunch. Most outlets had one or two.

What the yes and no camps actually spent

The propaganda was one side of the campaign. The money was another, and this part is public. Under the federal transparency rules, campaigns spending more than 50,000 francs must disclose their budgets to the Swiss Federal Audit Office. According to SRF, the yes side budgeted 3.84 million francs, of which more than 2 million came from Christoph Blocher. The no side budgeted 1.67 million, 1.2 million of it from economiesuisse.

That is the kind of fact a chatbot should be able to give you, and it is exactly the kind of fact that gets drowned out when one side of the internet is producing twenty articles a morning.

What to do with this

If you use a chatbot to understand a vote, ask it for sources and open them. An answer that leans on a domain you do not recognise, or on a telegra.ph page, deserves the same suspicion as a leaflet with no sender.

If you build with AI, the consequence is more concrete. Any system that retrieves from the open web will eventually pull in pages written to be retrieved. Deciding which sources a model may cite, and logging which ones it did, is not an optional extra in AI development. It is the part that decides whether the answer can be trusted. We make the same argument about where data goes in running models on your own infrastructure: the model is rarely the risk; the path to it is.

And if you only take one thing from this: the Neutrality Initiative was decided by Swiss voters, clearly. The pages written about it the next morning were written for the machines that will answer the next question.