The Reported-Not-Generated Standard: how to evaluate AI marketing content
AI marketing content has a specific quality problem that fluency masks: it sounds authoritative whether or not the underlying claims are true. A model that doesn't know your customer retention rate will generate a plausible-sounding one. A model that doesn't know which competitor feature you're better than will generate a confident-sounding comparison. The output reads fine. The claims are invented.
The Reported-Not-Generated Standard is a rubric that cuts through fluency to ask the right question about each specific claim: does this reflect something someone actually knew, observed, or decided — or did the model generate it on its own?
Claims that pass the standard are defensible. Claims that fail are a liability before they ship.
What "reported" means in this context
"Reported" is borrowed from journalism: a claim is reported when it traces back to something that actually happened, was observed, was measured, or was decided by someone with the standing to decide it. The model of a journalist who has a source for each assertion they publish is exactly right for marketing content, because marketing content makes claims in your brand's name.
In AI marketing workflows, sources take three forms:
Knowledge inputs. Facts provided to the AI in the brief, the brand memory layer, or the session context — founding dates, product capabilities, customer counts, positioning decisions. These are reportable because they came from the brand.
Cited evidence. Research, data, case studies, customer quotes — anything that was real before the model touched it. A statistic from a research report the model was given is reported. The same number the model "knows" from training data is not — you can't verify its source.
Explicit brand decisions. Positioning calls, terminology choices, the decision to name (or not name) a competitor — these are decisions someone made. When the brief documents the decision, the content that reflects it is reported. When the model infers a positioning without being given one, it's generated.
The standard in action: a claim audit
The most useful application of the Reported-Not-Generated Standard is a per-claim audit of a draft before it's approved. For each specific claim in the draft, ask: can I trace this back to a source in the inputs I provided?
| Claim type | Passes if | Fails if |
|---|---|---|
| A specific statistic or number | Traceable to data in the inputs | The model generated a plausible number |
| A product capability claim | Documented in the brief or product spec | The model inferred a capability |
| A competitive comparison | Based on a positioning decision in the inputs | The model assumed an advantage |
| A customer outcome or quote | From a real case study or customer interview | The model synthesized a plausible result |
| A market characterization | Based on research in the inputs | The model described the market from training data |
| A trend claim | Sourced from cited research | The model described a trend it inferred |
The pass condition is the same in every row: there is a traceable source in the inputs. The fail condition is also the same: the model filled in something that wasn't provided.
Why fluency makes this hard to catch
The standard exists because fluency hides the failure. A generated claim reads the same as a reported one. The sentence "67% of marketing teams saw measurable results within 90 days" looks exactly the same whether a researcher measured it or a model produced it. The only way to know is to check whether the source exists.
This is a different problem from tone or style. Voice drift is visible in the prose. Claim generation is invisible without an audit. See why AI marketing content sounds generic even from capable models — the voice issue is what you notice; the claim issue is what actually creates risk.
The audit step is the mechanism that makes the standard operational. Without it, the Reported-Not-Generated Standard is just a principle. With it, it's a quality gate.
The scoring rubric
Across a full draft, a Reported-Not-Generated audit produces a clear verdict:
All claims reported. The piece is auditable and defensible. Approve for use.
Minor generated claims (non-central, easy to fix or remove). Edit out or replace before approval. Don't ship with generated claims even when they're minor — pattern matters.
Structural generated claims (core argument relies on unsourced specifics). The draft needs a revision round, not copyediting. Return to the brief and fill in the missing inputs before the model re-drafts.
Generated framing (the positioning itself isn't from the brand). This is the highest-severity failure. The content isn't off-brand in minor ways — it's describing a brand that doesn't match the actual positioning. Return to the brief, verify the Brand Memory Layer is loaded, re-draft.
In practice, the distribution matters. An 800-word blog post with two fixable minor claims is a normal editing task. A piece where half the specific claims have no traceable source is a drafting-process problem: the inputs aren't there, and editing won't fix it.
What this standard demands upstream
Applying the Reported-Not-Generated Standard is a downstream test, but it reveals upstream gaps. The most common:
Thin briefs. A brief that gives the model a topic and a format but no specific inputs will produce a piece that sounds authoritative and is mostly generated. The standard will flag it every time. The fix is better brief inputs, not better editing.
Missing Brand Memory Layer. Without specific brand facts loaded into the session, the model fills in plausible ones. The standard catches the gaps; the Brand Memory Layer closes them.
No cited evidence. When the brief includes data, case studies, or research, the model can draw on them. Without them, it draws on training data — which may be real but isn't verifiable in context. Build the habit of attaching source material to briefs.
Positioning underspecified. If the brief doesn't tell the model what position to take, the model takes one. It will be a plausible one. It may or may not match where the brand actually stands. The fix is documented positioning decisions in the inputs, not post-hoc editing.
The connection to evaluating AI marketing output
The Reported-Not-Generated Standard is one dimension of evaluating AI marketing output quality. It's the most important dimension for risk — it catches errors that are invisible at the surface — but it's not the complete picture.
A piece can pass the Reported-Not-Generated Standard and still fail on voice, structure, or audience fit. A complete quality rubric checks all dimensions. But those other dimensions are visible; this one requires deliberate checking because the failure mode is designed to blend in. See how to measure AI content quality for the full four-dimension rubric.
The Copper Sun approach to content quality is built around this standard: every specific claim should be traceable to a source, and any claim that can't be traced gets cut or replaced, not polished.
Frequently Asked Questions
Is the Reported-Not-Generated Standard only for AI content?
No — it's a useful quality test for any content, AI-assisted or not. The difference with AI is that generated claims are more likely and harder to spot because the output is fluent. Human writers also sometimes insert statistics they half-remember or comparisons they inferred; the standard catches those too. AI just makes the problem more systematic and harder to catch by reading alone.
Does this mean AI can only write about things already documented?
Yes, for specific factual claims. For analysis, synthesis, explanation, and structure, the model can do work that isn't just re-reporting inputs — it can find patterns, draw connections, explain concepts clearly. But any specific claim that carries evidentiary weight — a number, a capability, an outcome — needs a source in the inputs. The standard doesn't limit what AI does; it defines what has to be verified before use.
How do I handle a claim I know is true but can't source quickly?
Cut it. If you can't verify it quickly, you can't verify it before it ships. Either find the source and add it to the inputs for a re-draft, or remove the claim. "I'm pretty sure this is right" is not the Reported-Not-Generated Standard — it's hoping. Marketing content under your brand's name deserves a higher bar.
What tools help apply this standard?
No tool does it automatically — the standard requires knowing what sources you provided to the AI and which claims in the output are traceable to them. The practical tool is a per-claim check during editing: for each specific fact or comparison, ask whether you can point to its source in the inputs. That check takes five to ten minutes for a typical blog post and is the highest-value quality step in an AI marketing workflow.