Can Brute-Force Tactics Influence an LLM's Political Opinions?

I've recently seen well known PR firms selling political influence campaigns built around networks of AI-generated "stealth sites" meant to sway the output of ChatGPT and other large language models (LLMs). They're called stealth sites because the content is entirely AI-generated and was never meant to rank in Google. Their only job is to be read by AI crawlers and, the theory goes, to tilt what AI tells voters. The pitch to clients comes with a second selling point: because the sites are hard

The pitch to clients comes with a second selling point: because the sites are hard to detect, they can influence AI right under the opposition's nose, and no one will be the wiser.

The short version: you can't spam your way to a conclusion. On any well-covered issue, this strategy is at best ineffective and at worst a waste of budget, and the "no one will notice" promise is the riskiest part of the pitch. Here's why it doesn't work, where the real exceptions are, and what I recommend instead.

The Stealth Site Playbook

During a recent consulting engagement, our team reviewed a proposal from a well-known global communications firm that promised to change how LLMs talk about the client's signature issue. The plan was ambitious and expensive, and it rested on a misunderstanding of how LLMs work.

The firm proposed to:

Register a batch of new domains and present them as generic local news sites (e.g., SouthDallasDailyTimes.com).
Fill them with AI-generated content: long thought pieces, rewritten news and fabricated surveys favoring the client's position along with articles attacking opposing views. The mix was meant to influence AI answers across a broad range of questions.
Ignore Google entirely. The sites' only job was to be read by AI.
Publish at scale: hundreds of articles to start, with the expectation that the network would grow so large that ChatGPT and other LLMs would be forced to treat it as a credible source and adjust their conclusions accordingly.
Invent authors and social media accounts to "add" a layer of credibility to LLMs beyond the content itself

The pitch was confident that this brute-force approach would quickly displace the authoritative sources already covering the issue.

This isn't a one-off. Similar strategies were used widely during the 2026 US election cycle. RankBee's software identified multiple networks of this kind and in our tracking, these sites were rarely, if ever, cited.

Why the Logic Sounds Reasonable

The pitch draws on an idea most people have heard: an AI model starts as a blank slate, and the data it takes in determines what it concludes. If that's true, why couldn't thousands of new articles arguing one side of a debate shift the model's view? Wouldn't the AI eventually decide its earlier conclusion was wrong?

The flaw is that this picture assumes LLMs can't tell sources apart. It assumes a model gives an article on a brand-new domain with no history the same weight as reporting from Reuters, Forbes or The New York Times.

The reality is that credibility is a nuanced, weighted and measured variable not based strictly on raw article and post counts.

Let me separate the two ways content can reach an AI's answer and how the stealth site strategy fails at each one for different reasons.

Failure Point 1: Live search.

The sites never show up.

When ChatGPT, Perplexity, Gemini or Google's AI Overviews answer a question about a current issue, they usually run a web search first and build the answer from what comes back. Those searches rely on search engine indexes and rankings: Google, Bing and similar.

That's the fatal flaw in the pitch. A site deliberately built not to rank in search engines is, for the most part, a site AI search won't retrieve. The strategy's defining feature, staying out of Google, is the same thing that keeps it out of AI answers.

And even when an unknown site does get retrieved, AI assistants look for confirmation across trusted sources, not repetition across unknown ones. Ten obscure sites saying the same thing don't outweigh one credible outlet saying otherwise.

Failure Point 2: Training data.

Too slow, Too Filtered.

The other route and the one that is critical to the strategy working is getting the information into a model's training data. Here two critical issues immediately appear:

Timing: Models are trained on snapshots of the web and updated on cycles measured in months. Content published during a campaign may not reach a model until well after Election Day rendering the entire engagement useless.
Filtering: AI developers filter and deduplicate training data, and low-quality, repetitive, machine-generated text is exactly what those filters are designed to catch. The content on these sites is fully AI generated and has little to no editorial oversight so it would be difficult for them to make it through advanced filtering.

Failure Point 3: "No One Will be the Wiser"

Everything is found eventually.

The pitch behind stealth sites treats secrecy as a feature. In practice, it's the biggest risk.

These networks get found. Groups like NewsGuard, the American Sunlight Project and whole groups of academic researchers and journalists, actively hunt for them. Fake authors, fake social accounts and invented surveys are exactly what platforms classify as coordinated inauthentic behavior and remove.

For a campaign, exposure becomes the story. A headline that says a campaign ran fake news sites to manipulate AI will cost far more than the program ever could have gained. And depending on the jurisdiction, fabricated polling and undisclosed political content may raise legal and disclosure questions.

We would never recommend this approach and we advise organizations and campaigns talk to counsel before signing off on anything that looks like this.

But Pravda and Other Groups Prove It Works ...

Over the last few years, you've probably seen headlines suggesting the opposite of what I just said. However there is fact to separate from fiction so let's go over the most commonly cited "network" of influence.

The Pravda Network

Pravda is a Moscow-based network of roughly 150 domains and is estimated to have published 3.6 million political articles in 2024 alone in over 40 languages, in a tactic researchers have dubbed "LLM grooming." In a March 2025 audit, NewsGuard found that ten leading chatbots repeated the network's false narratives about a third of the time.

However a closer look shows their audit tested prompts built around the network's own narratives, and many were worded to lead the chatbot toward the false claim. Independent testing published by Al Jazeera using broader prompts found false claims only about 5% of the time, and most references to Pravda sites were debunking them. Where the network did break through, it was mostly on obscure questions where little other coverage existed and this was a state-funded operation publishing millions of articles over several years, far beyond any campaign budget, that was publicly exposed by name.

The "250 Documents" Study.

In October 2025, Anthropic, the UK AI Security Institute and the Alan Turing Institute prepared a joint report concluding that as few as 250 documents slipped into training data could plant a hidden "backdoor" in a model. That sounds alarming, but the experiment taught models to output gibberish when they saw a specific trigger word. That's a security finding, not evidence that you can change what a model concludes about a widely debated policy issue.

The Real Exception: Data Voids

There is one place where stealth sites can partially work: topics with little authoritative coverage. Think of a rural county ballot measure, small town mayor/city council elections, a down-ballot race (local courts for example in Texas), or a very newly coined policy term.

When there's almost nothing credible to retrieve, an AI may cite whatever exists, including low-quality sites. This is where the Pravda network got most of its traction.

The lesson isn't "buy a stealth network." It's this: if your issue lives in a data void, fill it first with legitimate, attributable content and earned coverage before someone else fills it for you. Then monitor it, because your opponents can try the same tactic.

What AI Actually Treats as Credible

So why won't the stealth site approach work? Because these sites have none of the signals that tell an AI a source is worth trusting:

Independent Mentions: other credible sites referencing, linking to and talking about the source.
Real Authors: named writers with real bios and a real track record
Entity Connections: clear links between an author, the organizations they're tied to and the subjects they're known for, so that when they write on a topic, it fits who they are.
Corroboration: when an AI assistant retrieves sources to answer a question, it looks for confirmation across trusted sources, not repetition across unknown ones.

Spinning up websites with fake authors, invented bios and a flood of content provides none of these. That approach may have worked in a narrow window early in AI search. It doesn't work now.

The Real Cost for Campaigns and Ballot Measures

This outdated thinking drove real financial and strategic decisions in 2026 across political parties, ballot propositions and other high-stakes influence campaigns. Stakeholders paid for a strategy built on a wrong idea of how AI decides what to believe, and took on reputational risk in the process.

You can't push millions of pieces of content at an AI and expect it to flip its view.

Moving the Needle Without Wasting Money

AI draws on a rich set of sources to answer politically sensitive questions. Having the right message isn't enough; it has to reach AI through the sources AI already trusts.

What Rankbee does:

Find the sources AI actually trusts. Our technology validates the sources that ChatGPT, Gemini and Google AI Overviews rely on most for your issue, candidate or election.
Show how those sources describe you among audiences. We map how trusted sources currently frame your policies, your positioning and the debate around them, and where AI answers are inaccurate, incomplete or tilted. We break down answers down by audience and specific geographies using our proprietary digital twin voters
Structure your message for humans and AI. We write and revise key messaging content to be clear and effective to humans and AI. The goal is that when AI answers questions about your issue, your position is represented accurately and in your own words.

If you're planning AI visibility work for a campaign, PAC, party or ballot measure, or you've been pitched a stealth site strategy and want a second opinion, we'd be glad to help. Book a campaign strategy call with our team.