Manipulating AI sources: Anthropic rules it out in writing

What exactly does the new policy prohibit?

Manipulating AI sources now has a written definition. Anthropic published an update to its usage policy on 8 October 2026, effective from 12 November. Among the prohibited behaviours is one not previously worded this way, and it speaks directly to the marketing trade.

The text prohibits «manipulating the sources from which search engines or AI systems draw answers by seeding them with content that misrepresents its origin, authorship, or independence». It adds an example that leaves little room: «networks of sites posing as unaffiliated sources corroborating the same claims».

The scope is worth reading carefully. This is a usage policy for Claude. It governs what may be done with the tool, not how anybody ranks inside it. Somebody running a network of sites without using Claude breaks no part of this policy.

Why does it matter if it only governs usage?

Because retrieval systems learn from the same patterns their makers write policy about. Conduct prohibited in writing today tends to become a signal the model learns to discount tomorrow.

There is a close precedent. Google spent years describing practices in its spam policies before penalising them automatically. Its site reputation abuse policy arrived before the algorithmic treatment of the problem, as we covered when analysing the latest spam update.

So the move is interesting even if no sanction reaches you. It states what the maker of the assistant you want citing you treats as manipulation.

What changes for anyone publishing content

For a company publishing on its own site under its own name, nothing. The prohibition describes the opposite extreme: content that hides where it came from.

Where the calculation does change is in three practices sold as normal in the industry.

1. Networks of apparently independent blogs linking to and citing the same client.
2. Sponsored articles published without saying they are sponsored.
3. Reviews and comparisons on sites that look neutral and belong to the manufacturer.

All three work today in traditional search engines. All three are exactly what this text describes.

What is worth doing now

What ages best is the dullest thing: making it clear who signs and who pays.

State authorship on every page, with a real person behind it. Label sponsored content as sponsored. And when working with publishers or collaborators, ask for the relationship to be visible. This is also what makes a source citable. An assistant that cannot attribute a claim to somebody tends not to use it.

In practice that means author and organisation markup done properly, part of the work on structured data. And a visibility strategy that does not depend on faking independence, as we set out in SEO for AI search engines.

What this policy does not say

It does not say how Claude picks its sources. It describes no selection or citation mechanism. Anybody presenting this document as a guide to ranking in Claude is stretching the text.

Nor does it announce automatic sanctions for websites. It is a product usage policy, with the usual consequence: losing access to the tool.

The value sits in the statement rather than the threat. It is the first time a maker of assistants has described in writing what it treats as contaminating the source, and that marks out the playing field for the coming years.

For an ordinary company the conclusion is reassuring and unglamorous. Publishing under your own name, saying who pays and adding something of your own will keep working, and it competes against less noise every year.

Frequently asked questions