How claims stay accurate

The Marketing Agent ties each claim in a draft to a source it read in that session, and checks every blog post and every draft that names your product in a separate pass before you see it. These checks make an unsupported claim less likely, but your review is still the last check.

Drafts start from your sources

The agent researches before it writes, and it writes from what it read in that session:

• Every claim about your product cites a source the agent read. A claim with no source is cut, not softened. The agent never invents a feature, a number, a quote, or a customer name, and never says that something shipped, is generally available, or is not behind a flag unless a source says so.
• Evidence is labeled. Each finding is an official statement, a measured result, a practitioner’s report, or the agent’s own inference, which is never written as a measured result. Numbers, quotes, and code keep the source’s exact words, and a search-result snippet is never evidence.
• Your own pages are compared. For a price, plan, limit, integration, security, or availability fact about your product, research checks your published pages and flags the one to fix if two of them disagree.
• Links and code are traceable. A blog post links only to sources the agent read, inline on the words each one supports. Code and commands appear only when copied from a file the agent opened.
• Private sources stay private. A draft never links to or names a private document. It describes one in words a reader could find, such as “the release notes”. In the chat, the agent cites it by path or link.

Without research findings, the agent stops and asks for sources. It does not draft a blog post from memory. For the sources it reads, see Marketing Agent.

A separate claims check

After drafting, every blog post and every draft that names your product goes through a claims check. The check reopens each source it can and looks at each image.

A claim is anything that says your product does, has, ships, supports, or measures something; any number; any named customer, company, or person; any quote; any date or status (beta, generally available, behind a flag); and any comparison with a competitor or with how people work today. Each claim gets one of four outcomes:

Outcome	Meaning	What happens to it
sourced	A source states it, with the same scope.	It stays.
unsourced	Nothing the check can read states it.	It is cut, or rewritten without the number.
contradicted	A source says otherwise.	It is rewritten to match the source, or cut.
overstated	A source supports a narrower claim: for example, the draft drops a date, a sample, or a plan, or states one result as general.	It is narrowed to what the source shows.

The check also flags a link to a page outside the draft’s sources, and a diagram that differs from the text.

Everything the check flags is fixed before you see the draft. For a blog post, if the fixes changed the draft, the check runs again on the final version and anything still open is reported to you. A blog post also gets a separate voice review, whose fixes keep each fact’s sourced wording.

To run the check again after you edit, click Claims check under the message box and send the message.

What you review before anything ships

1. The chat summary. It opens with the agent’s recommendation, then lists what it read, with sources, what it could not read, what it drafted, which images it added, any code or commands in the draft it did not run, and the assumptions it made.
2. The draft, open in the editor, where you can edit it directly. When a later turn changes a draft, the editor shows the changes with Accept All and Reject All.
3. The publishing step. The agent publishes nothing unless you ask and name the destination; otherwise you copy or download the draft and publish it yourself. A draft added to a documentation repository arrives as a pull request your team reviews and merges.

FAQ

Can a draft still contain a wrong claim?

Yes. The check compares the draft with the sources the agent could read, and a source can be out of date or wrong. Read the summary’s list of what the agent could not read, and check any claim that rests on one of those.

Does the agent make up a customer quote or a statistic?

It should not. A number, quote, or customer name needs a source the agent read, and the claims check marks one without a source as unsourced, which is then cut or rewritten. Before you publish, check each quote and statistic against the sources in the chat summary.

Does a LinkedIn post or newsletter get checked too?

Yes, when it names your product. A draft that does not name your product skips the check.