On Saturday afternoon a 12-page English site passed its build with zero errors. The copy check still flagged nine items. Nine indexable pages had no blocking issue, three new pages scored 100 each, and the total was 98.5.
A copy check scans body text against a fixed word list and fixed sentence patterns. Every hit returns the page, the line, the matched phrase, a suggestion and a penalty. The same draft scores the same twice. It says nothing about whether the content is correct, and nothing about ranking.
Why an AI detector is not the gate
We tried one. A thermostat spec page with short sentences and a parameter table came back at 87% AI probability. People write those pages the same way: many numbers, few connectives.
False positives block normal technical pages. False negatives let real filler through. Scores also move between models, so the result is not reproducible.
The gate is a deterministic rule set. The tradeoff: it catches high-confidence patterns only. Judgement about whether a claim holds still needs a person.
Nine rules
The list came from manual annotation of English B2B pages, ordered by frequency.
| Category | Examples | Fix |
|---|---|---|
| Empty opener | in today's, when it comes to | delete, start with a fact |
| Inflated meaning | stands as, a testament to, pivotal | replace with a checkable sentence |
| Sales words | cutting-edge, game-changer, seamless | replace with a number or an action |
| Filler | in order to, due to the fact that | compress to one word |
| Chatbot residue | I hope this helps, let me know if | delete the sentence |
| Empty ending | the future looks bright | end on the next step |
| Negative parallel | not only ... but also | split into two statements |
| Template shape | bold heading lists, emoji, curly quotes | plain paragraphs |
| Em-dashes | more than 6 per 1,000 words | split the sentence, use a comma |
A site can add its own forbidden list. If a brand never calls itself an "OEM partner", that phrase goes into the site config. Extra words affect that site only.
Scoring and what blocks a release
Each hit carries a weight, each category is capped, and the floor is 0. Grades run A at 90 and above, B from 80 to 89, C from 70 to 79, D below 70.
Blocking depends on page state. Indexable and approved pages cannot ship with an error. Draft and noindex pages are recorded without blocking, so a new page can hold its URL while assets are collected.
A full check runs seven stages: contracts, pages, assets, seo, facts, voice, similarity. Copy runs after facts, because a page with unsettled facts gets rewritten anyway. The similarity threshold is 28%; the highest score here was 0.42%.
Rewrite safety
The model may change wording, not facts. Five checks run before a write.
- numbers and units match item by item
- URLs and emails match
- code blocks and headings match
- link targets match
- the new score is not lower
Any mismatch rejects the write and states the reason. Preview is the default. A flag writes to disk, and a timestamped backup comes first.
Comparison is string level, not semantic. One wrong character blocks the page, including a stray space inside a voltage value.
Measured
One page moved from 94 to 100 in 71 seconds, on a local 27B model, with no API cost. Long pages run better one at a time.
voice 0 error / 9 warning 9 indexable pages
similarity 0 error / 0 warning max 0.42% / threshold 28%
pages 0 error / 0 warning 12 pages / 9 templates
Nine warnings all carry a page and a line. Two examples: a process page with 9 em-dashes in 658 words, over the limit of 6 per 1,000; two legal pages using curly quotes.
Five steps today
- Run the check without edits. Pick one English site and get scores plus line numbers.
- Start with the lowest page and change one page at a time.
- Copy numbers and units into a list before editing, then compare after.
- Rerun. If the score does not rise, restore the backup.
- Gate indexable pages. Record drafts only.
Where it fits
English B2B sites, 10 to 50 pages, stable product parameters. Teams of one or two people with no dedicated copy reviewer.
It misfires on consumer landing pages where short spoken sentences are the point. A Chinese site needs its own word list; the sentence patterns transfer, the vocabulary does not.
FAQ
Is an AI detector reliable for website copy?
Not as a gate. Technical pages with short sentences score high. Use it for observation, not for release decisions.
What score is publishable?
85 is the floor here, 90 is comfortable. Read the hits first. The same category repeating points to a writing habit, so change the structure rather than one word.
Can the local model change a spec by accident?
It cannot write without passing the five checks. Numbers, units, links, code and headings are compared item by item, and a mismatch rejects the write.
Can a Chinese site reuse the rules?
The word list changes. English rules catch openers like "in today's"; Chinese rules catch contrast pairs and dense enumeration. Patterns transfer, vocabulary does not.
How long does a run take?
Under a minute for 12 pages. A rewrite takes about 70 seconds per page, on a local machine, with no API bill.
Last updated: 2026-10-26