Agencies that build clusters of seemingly independent sites to give AI answers something to cite have a new exposure if Claude writes the content. Anthropic published an updated usage policy on October 8, effective November 12, and it addresses that tactic by name.
The policy adds a section titled “Do Not Engage in Deceptive Campaigns or Artificial Activity.” Among its prohibited uses is this one: “Manipulate the sources from which search engines or AI systems draw answers by seeding them with content that misrepresents its origin, authorship, or independence (e.g., networks of sites posing as unaffiliated sources corroborating the same claims).” Other lines cover fake personas, fake outlets, and fake reviews.
Search Engine Journal, in a report by Matt G. Southern, points out that the clause lives in the policy text rather than in Anthropic’s announcement. Anthropic’s own post describes the new section as a consolidation, gathering rules previously split among four separate policy areas: elections, fraud, privacy, and disinformation. It also says the section covers political and commercial activity alike. The September 15, 2025 version of the policy made no mention of search engines or of AI-generated answers.
Anthropic tied the section to its September threat intelligence report. That report described about 70 sites presented as local news outlets, all tied to an advertising agency based in France. That operation produced at least 8,913 articles in about 20 languages, according to Search Engine Journal’s summary, and most drew little real engagement. Anthropic removed the account and banned the organization.
The case is a useful warning because the outcome is missing. The report does not say whether the sites gained rankings or showed up in AI-generated answers. A network that never worked was still enough to trigger a ban. Teams should not read this as an enforcement threshold that depends on results.
A second change deserves careful reading. Search Engine Journal observed that automated publishing, which the earlier policy placed on its high-risk list, is absent from the revised list. Anthropic does not explain the removal in its post, so the reason is unknown. It is a change to one list, not a blanket exemption. The deceptive-activity section applies regardless, and a site network that hides its common ownership fails on provenance, not on how the text was produced.
For agencies, the distinction to hold is between corroboration and fabrication. Placing real expert commentary, original data, or genuine partnerships on third-party sites that readers can verify is ordinary digital PR. Publishing the same claim across properties that present themselves as unrelated, when one operator controls them, is what the clause describes. The Claude account holder is the party exposed, and consequences include suspension or termination of access.
That exposure falls on the agency’s tooling, not only on the client’s rankings. Google enforces its spam rules against websites, which can lose visibility. Anthropic enforces against accounts. An agency that loses its Claude access mid-engagement has a delivery problem that no ranking recovery plan covers.
Anthropic is the only AI company covered here, and this report makes no claim about other providers’ terms. Teams should check each tool in their stack on its own.
Brands auditing vendors should ask three direct questions. Which domains does the vendor publish to, and who owns each one? Does any content appear on sites presented as editorially independent that the vendor in fact controls? Which AI tools draft and publish that content, and under what account?
Before November 12, brands should require written answers and ask agencies to list every property carrying client claims. Any network that depends on looking independent is a liability under this policy, whatever its ranking record.
Anthropic, “2026 Usage Policy update” (October 8, 2026), and Search Engine Journal, reported by Matt G. Southern (October 9, 2026).
