Anthropic bans Claude-powered content networks built to sway AI answers

Anthropic has rewritten Claude's usage policy to explicitly ban feeding search engines and AI systems with networks of sites that hide who is really behind them, publishing the update on October 8 and setting it to take effect November 12.

The change consolidates rules that used to sit scattered across Anthropic's election, fraud, privacy and disinformation sections into one new block, titled "Do Not Engage in Deceptive Campaigns or Artificial Activity." It bars using Claude to build fake personas, outlets or reviews, to run sites that pose as independent of one another while pushing the same claims, or to sell tools built for that kind of campaign. A comparison from Search Engine Journal of the old and new policy text found no equivalent clause in the version Anthropic published in September 2025.

What changed in the policy

The new section targets a specific tactic: pointing Claude at content meant to populate sites built to look unconnected from each other while repeating the same claims, so a search engine or an AI answer reads the repetition as independent corroboration rather than one coordinated source.

"Manipulate the sources from which search engines or AI systems draw answers by seeding them with content that misrepresents its origin, authorship, or independence" — Anthropic's Usage Policy

The update also removes automated content publishing from Anthropic's high-risk use-case list, which otherwise requires a qualified human reviewer and an AI disclosure before outputs reach the public. Anthropic's post does not explain the removal, and says its ban on deceptive campaigns has not loosened — an automated pipeline still breaks the rules if what it publishes hides where the content came from.

The case behind the rule

Anthropic's own September threat intelligence report supplies the example its new section targets. Investigators linked a cluster of sites posing as independent local newsrooms, in a case the report labels GTG-54002, to a single operator: LKM Company, an advertising firm headquartered in France.

Each site ran an invented journalist byline and its own paired account on X. Anthropic says the operator stood the sites up in one short burst, buying the domains from France over a single stretch of weeks, then hosting every property on shared infrastructure behind one account. Claude did two jobs inside the operation: writing original articles for the fake outlets, and rewriting real reporting from working journalists into politically slanted versions aimed at specific audiences, among them readers in the Democratic Republic of Congo, Brazil, France and the United States. Anthropic says it shut the network down before it built a real following — the report ranks the activity near the bottom of the severity scale it uses for influence operations, reserved for content that never spread beyond the network's own pages and accounts.

What this means for software marketers

Few SaaS marketing teams run anything close to a 70-site network. But the tactic Anthropic now names overlaps with something far more common: programmatic content and "AI visibility" pushes built to get a product cited inside ChatGPT, Gemini or AI Overviews. Growth teams running those programs in-house, and the SEO or GEO agencies many of them hire to run them instead, are the audience this policy actually reaches. A content program that uses Claude to populate several nominally separate sites with matching claims about the same product now sits inside a policy Anthropic can enforce by suspending or ending the account behind it, on top of whatever a search engine does to the sites.

That second layer is the real change. Google's own spam policy already treats attempts to manipulate generative AI responses in Search as spam, enforced against the sites. Anthropic's update adds an independent point of enforcement against the account that generated the content — so a team whose agency gets a domain penalized by Google could also lose Claude access entirely, with no search console to appeal to.

Where the industry already toes this line

Spreading the same proof point across many properties has been an ordinary distribution tactic for years: syndicated contributor posts under rotating bylines, sponsored microsites that don't name their backer, review placements bought through networks that never disclose the buyer. None of that is new. What's new is that the company whose model powers a large share of these programs has now named the pattern specifically and tied it to a documented enforcement case, rather than leaving it as an implicit reading of a broader fraud clause.

• List every property, owned, sponsored or agency-run, that repeats the same product claim under a different byline or brand name
• Confirm each of those properties discloses its real owner, sponsor or any AI tool used to produce the content
• Check that no outside vendor is selling a service built specifically to run this kind of multi-site seeding