AI-assisted content can be useful when it transforms verified source material into a page with a real reader job, original value and an accountable editor. It becomes sludge when the production target comes first and the evidence, angle and reason for publishing are reverse-engineered afterwards.
That makes AI content versus human content the wrong argument. A person can produce generic landfill slowly. A model can help a knowledgeable editor turn interviews, operational records, product knowledge and evidence into a strong draft quickly. The tool does not decide whether the page deserves to exist.
The uncomfortable moment arrives when the queue is full of fluent drafts and the monthly volume has already been sold. Every document looks nearly finished. The research needed to challenge its business claims was never scheduled. Release the batch and plausible assumptions inherit the company’s name. Return it and the editor appears to be blocking the efficiency promised by the system.
My first question is not Was AI used? It is:
Which evidence controls this page, what would disappear if we did not publish it, and who is allowed to stop it?
If nobody can answer, the draft does not need a more advanced prompt. It needs knowledge, or deletion.
TL;DR
- The tool does not decide whether a page deserves to exist. The publishing target does.
- Give the model a source pack before a word count. Silence is not permission to invent.
- Every draft must survive five deletion tests before anyone polishes the prose.
- Name the person who can refuse publication, or the release gate is theatre.
Fluent, accurate and pointless is still a failure
Content sludge does not need hallucinated statistics or broken grammar. The most scalable form is plausible, inoffensive and almost impossible to remember. It can fail because:
- it restates common knowledge without changing a decision;
- it turns one thin idea into several pages competing for the same job;
- it removes qualifications that made the source accurate;
- it presents generic synthesis as first-hand expertise;
- it gives every company in the category the same advice with different nouns;
- it creates maintenance obligations for facts nobody owns;
- it overwhelms editors with enough reasonable prose to make serious checking unlikely.
This is why a humanised rewrite is not a cure. Changing adjectives, adding fragments and dropping in a first-person sentence can make the page less robotic while leaving its intellectual value at zero. Artisan sludge is still sludge.
Google does not ban the tool, and it does not rescue the output
Google’s current guidance says generative AI can help with research and structuring original material. It also warns that using generative tools to create many pages without adding value may violate its scaled-content-abuse policy (Google Search guidance on generative AI content).
The scaled content abuse policy focuses on many unoriginal, low-value pages created primarily to manipulate rankings, regardless of whether they were produced by automation, humans or a mixture. Google lists generative production without added value, scraping or stitching sources without value, and keyword-filled low-sense pages among its examples. Its newer guidance for generative search encourages useful, reliable, non-commodity work and explicitly warns against creating separate pages for every possible query variation (Google’s guide to generative AI features).
That gives AI SEO a much cleaner boundary than detector scores and arguments about punctuation. The relevant questions are purpose, value, evidence and editorial control. A human approval click does not neutralise a production system built to manufacture search coverage from interchangeable pages.
Give the model a source pack before giving it a word count
Write 1,500 words about X is not a serious brief. It rewards completion of length while leaving the model to infer the business. Before generation, assemble a source pack that controls what the draft may claim:
| Input | What it controls | What absence causes |
|---|---|---|
| Page job | The question or decision the page must complete | The draft expands across the whole topic |
| Reader and situation | Which risk, knowledge and next action matter | A generic audience with generic pain points |
| Primary claim | The proposition the page is prepared to own | Several soft assertions and no spine |
| Verified facts | Names, offers, scope, dates, locations and constraints | Plausible business details are invented |
| Evidence | What permits each consequential statement | Sources become decoration |
| Original contribution | Experience, method, data, example or decision tool | Commodity synthesis |
| Prohibited claims | Results, guarantees or relationships that cannot be stated | Sales language outruns reality |
| Known limits | Conditions that change the answer | Caveats disappear or arrive too late |
| Accountable reviewer | Who can correct the substance | Review becomes a style check |
This does not require another form for the sake of process. A transcript, annotated report, product document and short editorial brief may contain everything. The rule is that silence is not treated as permission to invent.
If the original-contribution field is empty, stop. Interview somebody. Inspect a real project. Document the process. Build a comparison with disclosed criteria. Merge the topic into a stronger owner. Prompt evolution cannot manufacture lived evidence from an empty brief.
Use AI where transformation is the actual job
Generative systems are particularly useful when meaningful source material already exists and the production task is transformation.
| Editorial task | Useful AI role | Boundary |
|---|---|---|
| Expert interview to draft | Extract themes and propose an argument | Check the transcript and retain conditions |
| Brief-to-draft comparison | Find missing questions, repetition and unsupported sections | Flag the gap; do not invent its answer |
| Alternative openings | Recast approved facts around different reader tensions | Select for truth and decision value |
| Format conversion | Turn known material into a procedure or comparison | Reject structures the evidence cannot support |
| Primary-source summary | Create working notes and bounded explanations | Verify material claims against the source |
| Tone application | Improve rhythm and remove filler | Do not increase certainty or fabricate experience |
| Page maintenance | Compare current source records with existing copy | Let the accountable owner decide what changes |
I would not release unsupervised claims about law, medicine, finance, credentials, product performance, customer outcomes or personal experience. That is not because a person writes infallibly. It is because fluent generation does not supply the missing evidence or professional responsibility. AI can help express what the source permits. It does not decide what the business knows.
A human near the keyboard is not editorial control
Human reviewed means very little unless that person had the evidence, competence, time and authority to reject publication. Three responsibilities need owners, even when one person holds all three:
- Evidence ownership: somebody can show where facts, numbers and examples came from.
- Subject ownership: somebody can judge whether the method, limit and recommendation are sound.
- Editorial ownership: somebody can decide whether the page has a distinct job and earns publication.
A proofreader is not necessarily a subject reviewer. A subject specialist may not see cannibalisation across the site. A marketing manager cannot automatically approve a regulated or technical claim. Naming the responsibility prevents everybody assuming somebody else checked the dangerous part.
The reviewer needs three real decisions:
- Release: evidence, original value, page ownership and accountable review are complete.
- Return: the page job is valid, but named evidence or limitations are missing.
- Merge or kill: the page duplicates another asset or exists to satisfy output volume.
If the only available actions are approve and make it sound more human, the person is decorating the production line.
Make every proposed page survive five deletion tests
Run these before polishing the prose.
- Competitor swap
Could the article move to another company's site after changing brand and place names?
- Site deletion
If it vanished, which unique fact, method, example or decision would disappear?
- Source delta
What did the model receive that a generic prompt did not: interview, dataset, record, operational method or signable judgement?
- Byline subtraction
Remove the author's load-bearing passages. Does an equally useful generic explainer remain?
- Collision
Which existing page already owns the question, and why can this material not be absorbed there?
A readable, accurate, on-brand draft can fail all five. That is the problem conventional content QA misses.
Non-commodity output does not require a theatrical contrarian theory. A precise limitation, disclosed cost model, annotated example, diagnostic sequence, documented failure or honest refusal can carry more value than a proprietary-sounding framework with no evidence. The test is not whether the article sounds different. It is whether the reader can do, judge or understand something they could not obtain from the obvious synthesis.
Audit claims before editing sentences
Style editing cannot repair an unsupported proposition. Classify every material statement:
| Claim state | Public treatment |
|---|---|
| Verified fact | State it within the supported entity, date and scope |
| Business-supplied fact | Attribute or verify it where customers may rely on it |
| Customer report | Identify the reviewer or population and avoid universalising it |
| Professional judgement | Present it as the author’s rule, interpretation or preference |
| Inference | Name the evidence and calibrate certainty |
| Conflict | Show the disagreement or withhold a definitive statement |
| Unknown | Remove it or return it for verification outside public copy |
Numbers carry their own proof burden. A percentage needs its population, period, denominator, method and source. A price needs currency, scope, tax treatment and applicable date. A result needs a baseline and enough context to stop timing becoming a causal story.
Keep the limit beside the claim it controls. Smooth prose often places the promise first and the inconvenient caveat later. The reader should not have to remember a correction from six paragraphs away.
Sitting on a queue of fluent drafts nobody has been allowed to reject? We run the deletion tests on them and tell you which to release, return or merge.
Scale defensible assets, not generated words
Automation can improve parts of the process without deciding the publication target. Scale source extraction, terminology checks, duplicate detection, evidence records, claim comparison and QA when those operations improve consistency. Use models to challenge briefs, locate gaps and transform approved knowledge. Do not scale:
The safe production unit is not words generated. It is defensible asset released. That change sounds small and wrecks most volume-first pricing models. Good. A growth marketing system should reward pages that clarify decisions, support customer acquisition and survive scrutiny, not make the monthly report look industrious.
The release gate must be allowed to stop the queue
Before publication, require every draft to pass this gate:
One failed line does not always mean deletion. A valid page can return for evidence. A bloated draft can be reduced to its useful core. Two overlapping pages can be merged. A weak title can be narrowed when the body holds a defensible decision.
But no must remain possible. If every draft will publish because the subscription generated it, the release gate is theatre.
Name the person who can refuse publication
For the next proposed AI-assisted article, do three things before requesting prose:
- Name the reader decision the page will own.
- Assemble the evidence and original contribution the model may use.
- Name the person with authority to stop publication.
If one is missing, you do not have an article brief. You have a topic, a production target and a machine ready to sound helpful.
Stopping it is not a failure of AI marketing. It is the moment the editor proves they are more than the final approval click.
Good to know
FAQs
No. Google says generative AI can help with research and structuring original material. Its scaled-content-abuse policy targets many unoriginal, low-value pages created primarily to manipulate rankings, whether they were produced by automation, people or a mixture.
Not on its own. Changing adjectives, adding fragments and dropping in a first-person sentence can make a page less robotic while leaving its intellectual value at zero. Artisan sludge is still sludge.
A source pack: the page job, reader and situation, primary claim, verified facts, evidence, original contribution, prohibited claims, known limits and an accountable reviewer. If the original-contribution field is empty, stop and go and get knowledge.
Competitor swap, site deletion, source delta, byline subtraction and collision. A readable, accurate, on-brand draft can fail all five, which is the problem conventional content QA misses.
Next step
Want a release gate that can actually say no?
We build source packs, run the deletion tests and name the evidence, subject and editorial owners so AI-assisted pages earn publication instead of filling a quota.