Anthropic's new Claude policy bans fake site networks built to sway AI answers

Anthropic's rewritten Usage Policy for Claude now bans using its models to manipulate the sources that search engines and AI systems draw their answers from. PPC Land reports that the text effective September 15, 2025 had no comparable rule.

What the rule covers

The clause sits in a new section, Do Not Engage in Deceptive Campaigns or Artificial Activity. Its six bullets prohibit, among other things:

• seeding search and AI sources with content "that misrepresents its origin, authorship, or independence", with the example of "networks of sites posing as unaffiliated sources corroborating the same claims"

• fake personas, accounts and media outlets, with fake reviews now named next to astroturfing and sockpuppets

• distributing content through ostensibly independent sites to hide that it comes from a common source

• building tools for these campaigns, or marketing and selling them as a service

What it does not cover

The bullet does not mention optimisation or content volume. Structuring pages so AI systems can quote them, or publishing more of them, falls outside the wording. The line it draws is about deception: who wrote a page, and whether it is really independent.

It also governs production only. The policy restricts using Claude to make the material. Ranking stays with the engines, and Google clarified in May 2026 that its spam policies apply to AI Overviews and AI Mode. The policy binds Claude users, so content made with other models is outside it.

One gap stays open. Concealing sponsorship is banned when content aims to influence "public opinion, officials, or other decision-makers". The policy does not define decision-makers, and it does not say whether a purchasing decision counts.

Why vendors should read it closely

Two changes put the risk on sellers. The last bullet reaches anyone who markets these tactics as a service. And enforcement now starts on suspicion. The old text acted "if we learn that you have violated" the policy. The new one acts "if we suspect that you may have violated" it, and adds a warning and a limit to the list of responses.

The tactic is real. In July 2025, an agency showed that publishing the same list on 10 expired domains with a Domain Rating below 5 got ChatGPT and Perplexity repeating it. A package like that, built with Claude, is now the exact pattern the policy names.

What to do

If you buy AI visibility work, ask how each citation is earned. Pages on sites you own, under your own name, and coverage from outlets that are genuinely independent do not match this wording. A cluster of anonymous sites that all agree with each other does.