Three months after the campaign lands, your CMO opens the case study at an industry talk and asks where the bolded conversion-rate figure came from. You search the brief, the draft, the model chat history. The figure is plausible. The figure is also untraceable. The model wrote it; the editor kept it; the audit trail goes cold at the boundary between the chat tab and the CMS.
What “defensible” actually means
A piece of marketing output is defensible when, three months after it ships, you can answer four questions in under a minute:
- Source. Where did each non-trivial claim — every statistic, every named company, every quoted analyst — come from?
- Tier. What kind of source was it? A peer-reviewed paper, a regulator filing, a vendor blog post, a press release? The evidence standards name the tiers.
- Verification. Did a human open the source and confirm the figure or the quote?
- Authorship. Who in your team last touched the claim before publish?
A draft that cannot answer those four for every load-bearing claim is undefendable. It is also, at most companies, the default state of AI-augmented copy.
Why “the model wrote it” is not an answer
Three reasons the trail goes cold when the model is in the loop:
- Hallucination is invisible at the surface. The model produces a figure, a journal title, a quote attribution. The figure is plausible. The journal title is plausibly named. The quote sounds like a thing the analyst would say. None of those signals tell you the figure exists. Your editor reading without the source link cannot tell either.
- Chat history is not a citation. The model’s reply sometimes includes a URL, sometimes does not, and when it does the URL frequently 404s or points to a page that does not contain the claim. A screenshot of the chat is not a source; it is a record that the model said something, which is what you already knew.
- Your CMS does not know. The published post carries the claim and the byline. It does not carry the chain of custody from claim to source to verification. By the time anyone asks, the chain has to be rebuilt from memory.
The smallest discipline that works
You do not need a documents-management system. You need three things, applied to every load-bearing claim before publish:
- A linked source for every statistic, every named company, every quoted person. Not the model’s reply — the original document. If the original cannot be found in five minutes, the claim does not ship.
- A tier annotation. A simple suffix or footnote noting the tier the source sits in. Tier 1 sources need no qualifier in the copy; Tier 4 sources need a qualifier (“according to [vendor]‘s 2024 report”) so the reader can weight it themselves.
- A verification stamp. A name and a date in the working draft, sitting beside each load-bearing claim. The stamp does not appear in the published post; it appears in the editorial workflow that produces it.
That is the floor. The Playbook’s verified case-study register is the same discipline applied to one specific class of claim — named-company case studies — across the whole Playbook. The bar each entry has to clear sits on the case study standards page; the register itself reads inside the subscriber library. The pattern scales.
The trail belongs to the team, not the model
The model is a writing assistant. It is not a fact-checker. It is not a librarian. It is not your audit log. The team that treats it as any of those things ships output it cannot defend, and pays for that in the meeting where the CMO asks the question that has no answer.
The audit trail is the discipline that exists before the model is in the loop and survives after the post is published. The model speeds up the writing; the trail is what makes the writing land.
If you want the long-form treatment, the evidence standards are the public version of the audit floor, and the full Playbook is the rest of the operator’s discipline.