From fake quotes to fake sources: A Russian network’s attempt to have its own Wikipedia for Armenia

In April 2026, just two months before the parliamentary elections in Armenia, the Russian organization Social Design Agency (SDA) developed a “clone” similar to Wikipedia, designed to automatically create new pages. But its goal was not only to reach readers. According to a Bloomberg investigation, the system was also designed to influence Google’s search algorithms and artificial intelligence models, which gather information from content available on the Internet.

This became known from a Bloomberg investigation published on June 24, 2026 , based on 73 documents leaked from the SDA. Some of them were previously revealed by the Armenian Fact Investigation Platform. The documents show that the project on Armenia was part of a broader information activity of the SDA.

This case shows that today, in order to spread disinformation, it is not necessary to directly provide people with false information. One can also try to influence the sources from which people or artificial intelligence systems receive information. In this case, not only the reader is targeted, but also the information that may become available to them in the future.

What were they planning?

The SDA document, dated April 14, described a project for Armenia that the organization called a “self-populating knowledge base.” It was to be structured like Wikipedia and constantly updated with new pages.

The logic of the project was simple: to create information that would appear as ordinary reference content, but at the same time would be able to appear in search results and be accessible to systems that collect information from the Internet.

According to Bloomberg, the project document even described a specific method. The site was supposed to track which articles were most frequently searched for or read, and place relevant content, links, and information blocks on those pages. In other words, it was necessary not to simply create a large amount of material, but to deliver it to the pages that the average user was more likely to encounter.

The project’s authors also predicted quite specific results. The documents stated that in the first three months the site should have about 5,000 visitors per day, and the audience should subsequently grow by 30 percent per month. The numbers themselves do not mean that these indicators would be possible in reality, but they do show the scale of the calculated project.

The project for Armenia was also not the first such attempt by the SDA. A similar “knowledge base” was already in operation for Germany, and by the beginning of 2026, according to internal SDA documents, it already included more than 200 thousand pages. The base was planned to be continuously updated, including to maintain its visibility in search engines. Another goal of the project was to make this content available to artificial intelligence systems. The documents, in particular, mentioned the editing of 500 articles per month and the training of this material on six different AI platforms.

It is this parallelism that shows that the Armenian project was not a separate initiative, but rather an extension of the approach already used in Germany to another country. The only difference was in the target country and local content, while the method was the same: creating seemingly trustworthy information sources and using them to gain access to search engines and AI platforms.

What Bloomberg actually found

Bloomberg reporters didn’t stop at just examining the documents and tried to verify whether any of the projects described in them had actually been built. Using historical domain registration data, they discovered three websites: spyurk.cyou, sevan.info, and khachkar.info, all registered in January 2026 and hosted on the same Russian IP address. The sites contained content related to Armenia, much of which was simply copied from the Russian Wikipedia.

But the sites weren’t designed for regular visitors. They were built to automatically redirect visitors to other pages. According to a Bloomberg study, the sites’ primary function was to create content available to search engine bots—systems that scrape data from the web to power search results and AI models.

The project documents stated that the site was to be hosted in Turkey to disguise its Russian origin. In reality, spyurk.cyou was hosted by a Delaware-registered company with an office in Turkey. All three sites were terminated by their hosting provider on June 9, without a public explanation.

But what Bloomberg discovered is not limited to these three sites. The investigation also uncovered a wider network, including erevan.one and a number of other sites operated by the SNG-Media group, which cooperates with the SDA. Unlike the three sites mentioned above, these resources were located in Russia and were linked to the Russian telecommunications regulator, Roskomnadzor. They were not shut down on June 9. Thus, the three sites discovered, according to Bloomberg, were only part of a wider infrastructure.

At the same time, it is important to distinguish between the purpose described in the documents and their actual outcome. Bloomberg has not confirmed that this content actually appeared in the training data of any major AI company or significantly affected search engine results. What is confirmed is what was intended and what sites were built, but not what impact their content ultimately had on the relevant systems.

What makes this approach different?

It is important to understand how this approach differs from earlier methods of spreading disinformation. In the case of classic trolling, the impact occurred immediately, through a fake profile, post, or comment. The target was a specific person, and that content could be countered or refuted on the same platform.

In this approach, the target of influence changes: it is not just about directly influencing the reader. There is also an attempt to influence search engines and language models that find and present information to the user. Ekaterina Sedova, a researcher at the Atlantic Council , describes this tactic as an attempt to flood search engines with artificially large amounts of content that links to each other, so that the algorithm perceives its multiple repetition as a signal of trustworthiness.

The problem is that while a fake post on a social network can be found, flagged, and refuted, the source of the material in the language model’s training data may remain virtually inaccessible to the user. The person receiving a response from a chatbot often has no idea at all what source the specific claim came from.

Sedova also points out the vulnerability that can arise in languages ​​like Armenian, which have a relatively limited amount of online content. In the case of AI systems working with English, there are greater resources and professional capabilities to detect such manipulations, while in the case of smaller languages, the lack of relevant specialists and data can make it difficult to detect such content. In the case of Armenian, this is especially important, since the volume of online content in Armenian is significantly smaller than in English. Therefore, even a relatively small amount of artificially created content can occupy a more noticeable place in the available information.

Conclusion

Jacob Rogers, legal counsel for the Wikimedia Foundation, commented on the Bloomberg revelations, expressing doubt that such clones can have long-term success without a large and active community of editors. “We don’t see such clones having long-term success at all. They can reproduce Wikipedia content for a while, but they can’t update it regularly,” he said.

This problem is also visible in Armenia. The three websites discovered by Bloomberg only existed for a few months before their hosting provider stopped serving them without a public explanation.

However, the closure of these three sites does not mean that the experiment has completely failed. Bloomberg also discovered other sites within the same network that were not closed on June 9. In addition, a larger version of the same approach was developed for Germany. These facts show that the Armenian project was not an isolated initiative, but part of a broader strategy, the aim of which was not only to spread individual false information, but also to influence information search and presentation systems through artificially created content.

Therefore, this story is important not only as a short-lived and technically unsuccessful experiment with three sites. More important is the method that can be used again, with less obvious errors and more difficult to detect traces. Shutting down a particular site may eliminate part of the infrastructure in question, but it does not in itself eliminate the approach underlying such actions.

Hripsime Hovhannisyan

The material was prepared within the framework of the Disinformation Flows Monitoring Project.

#CivilSocietyCooperation #ishrarmenia #ishrfactcheck

The project is implemented with the financial support of the Federal Foreign Office of the Federal Republic of Germany.