By KYN AI Advisory Team — AI implementation specialists, Singapore
AI-drafted content holds up when it passes three tests before publication: a low and falling edit rate, a fact check that traces every figure to a source, and a scan that rejects filler phrases. Rankings depend on the page being useful, accurate and specific. These three measures catch the main ways AI drafts fail. They are cheap to run and most can be scripted.
What is an edit rate and why track it?
The edit rate is the share of a draft a person changes before approving it. Measure it as lines changed over lines in the draft, or words changed over words, and stay with one method.
It tells you where the problem is:
- High and steady. The brief or the style sheet is wrong. Fix the instructions, not the reviewer.
- High and rising. Something changed: a new model, a new topic area, a looser brief.
- Low and steady. Either the process works or the reviewer has stopped reading. Sample to find out.
- Low on some topics, high on others. The agent lacks source material on the high ones.
Record the edit rate per article and per batch. Do not set a target on day one. Take a baseline of 20 articles, then watch the direction. A falling rate with stable review time is the sign the system is learning your rules through better briefs.
How do you fact check an AI draft?
Check claims in a fixed order, and make the agent help rather than judge.
- List every checkable claim. Numbers, dates, names, statutory rules, prices, quotes. The agent extracts them into a table.
- Attach a source to each. An official page, a regulator, a vendor's own documentation. If the agent cannot attach one, the claim is flagged.
- Open the source. A person or a script confirms the page says what the draft says. Models can cite pages that do not contain the claim.
- Delete what cannot be confirmed. Leave the figure out. An article without a number is better than an article with a wrong one.
- Date it. Write the check date on the article.
Rules that help: any percentage, multiplier, price, deadline or "within N days" about the outside world needs an inline link to an authoritative source. Examples of authoritative: the tax authority, the registry, the vendor's pricing page. A blog post summarising those does not count.
Never allow invented client names, statistics, quotes or results. Describe clients by industry only. If the example is illustrative, say so.
What are banned phrases and how do you scan for them?
Banned phrases are words that add no information or make claims you cannot prove. Typical groups:
| Group | Examples | Why it fails | |---|---|---| | Superlatives | unproven ranking claims | Unprovable | | Hype | adjectives with no measure | No content | | Promises | assurances of outcomes | Cannot be kept | | Filler openers | "In today's fast-paced world" | Delays the answer | | Punctuation tics | exclamation marks, em dashes | Read as machine-written |
A script scans the draft against a list and returns the line numbers. It takes seconds and it never gets tired. Keep the list in one file, version it, and add to it when a reviewer removes the same word twice.
A scan only catches words. It does not catch vague sentences that use clean words. Add a second check: every paragraph should contain a number, a name, a step or a definition. If it contains none, it is filler.
Does Google penalise AI content?
Google states that using AI to generate content is not against its guidelines. What it targets is content made mainly to manipulate rankings, with little value for readers. Read its guidance on generative AI content and its people-first content guidance.
The practical reading: the production method is not the test. Accuracy, original detail and usefulness are. A page with a real table, a verified figure and a direct answer in the first paragraph meets that test regardless of who typed the first draft.
What does the checklist look like in one place?
Before sign-off, a page passes when:
- the first paragraph answers the question in two to four sentences;
- headings are questions a reader would ask, and the count is within the template;
- every external figure has an inline source link that opens;
- no banned word appears;
- every paragraph has a number, name, step or definition;
- client examples name an industry, not a company;
- the title and meta description are within length limits and state the answer;
- the edit rate for this article is recorded;
- a named person has read it end to end.
That last line is not a formality. The scripts catch form. A person catches the paragraph that is fluent and wrong.
How do you stop quality sliding over time?
Re-read a sample of approved articles one month later. Refresh stale facts in place rather than deleting the page, since deleting a live page loses its history. See why you should never delete a blog article. For the production side, see the content agent workflow for a 200-article plan.
What does an edit-rate log look like?
A log needs four columns: article, batch, words in the draft and words changed. The numbers below are illustrative.
| Batch | Articles | Average edit rate | Reviewer minutes per article | |---|---|---|---| | 1 | 20 | 34% | 22 | | 2 | 20 | 27% | 17 | | 3 | 20 | 18% | 12 |
A falling rate with falling review time is the pattern you want. If the rate falls but review time rises, the reviewer is fixing things the edit count does not capture, such as moving paragraphs. If the rate is flat across three batches, read the edits themselves. Group them: wrong facts, wrong tone, filler, structure. The largest group tells you which instruction to change.
What counts as a filler paragraph?
Filler is a paragraph that a reader could delete without losing information. Test every paragraph with three questions:
- Does it contain a number, a name, a step or a definition?
- Could it appear unchanged in an article on a different topic?
- Does the next paragraph make sense without it?
If the answers are no, yes, yes, cut it or rewrite it with details from this topic. The second question is the strongest. Generic paragraphs repeat across articles in a batch, so a script that flags near-identical paragraphs between drafts catches them cheaply.
Which fact errors do AI drafts make most?
Check these first, since they recur.
- Out-of-date rates and deadlines. The draft states a figure that was correct two years ago. Confirm on the regulator's current page.
- Right number, wrong scope. A rule that applies to listed companies is described as applying to all.
- Merged sources. Two documents are blended into one claim that neither makes.
- Invented specifics. A named study, a client quote or a percentage with no origin. Delete it.
- Wrong jurisdiction. A Singapore rule applied to a Hong Kong reader.
A short pass for these five takes about five minutes per article and catches most serious errors.
Where KYN fits
KYN's article engine runs the scripted checks above before any draft reaches a person, and publishing waits for a named person. The agent has its own restricted account and every action is logged with the actor. The system runs in a cloud account in the client's name. The engine costs about S$0.06 per article for the draft, excluding review time.
Engagement starts with 14 days of review: we inventory the tools your writers and reviewers use today and interview the people who edit drafts. The full build, including the checks and the edit-rate log, ships in under 90 days. See marketing systems.
Frequently asked questions
What is a good edit rate?
There is no universal number. Take a baseline from your first 20 articles and aim for a steady fall. If the rate drops to near zero, sample the output, since the reviewer may have stopped reading closely.
Should I disclose that a page was drafted with AI?
Disclose where your readers or your sector expect it. Either way, a named person is accountable for the content, and that is the statement that matters to a reader.
Can a script fact check an article?
It can check that links open, that a number in the text appears on the source page, and that required links exist. It cannot judge whether the source supports the claim in context. A person does that last step.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
