A language model rarely copies a paragraph whole. That is what makes checking harder, not easier. The typical failure is a sentence that no scanner flags because it shares no long run of words with any single source, yet sits close to how three different pages already phrase the same point. Verbatim overlap is the rare case; near-duplicate phrasing is the everyday one.
There is also the risk nobody checks for: your own site. Publishing a second page that says what an existing page already says splits whatever authority both had, and the second page usually ranks for neither.
What the scanners catch, and what they do not
| Check | Catches | Misses | Cost |
|---|---|---|---|
| Plagiarism scanner | Verbatim passages, long shared runs | Paraphrase that keeps structure and order | Paid, per check |
| Search engine phrase test | Phrasing already common on the web | Paywalled or unindexed sources | Free |
| Site-internal search | Self-duplication across your pages | Overlap in intent rather than wording | Free |
| AI-detector score | Nothing reliably | Almost everything | Free to paid |
| Human read for provenance | Unsourced claims, invented figures | Slow, does not scale | Time |
AI-detector scores should not be used as evidence. They produce false positives on ordinary human writing, they are inconsistent between tools, and they vary with text length. If a client or a platform requires one, treat the result as a formality to satisfy rather than a finding to act on.
The five-step check
1. Scan for verbatim overlap
Run the draft through a plagiarism scanner before editing anything else. Look at each flagged passage and decide: quote it with attribution, rewrite it, or cut it. Read the matched source rather than trusting the percentage — a match on a phrase like "small business owners" is noise, while a match on a distinctive sentence is not.
2. Test your distinctive phrases in a search engine
Take the three sentences in the draft that carry the actual point, put each in quotes, and search. If the exact sentence already appears on the web, the draft did not originate it — and if it appears on a page you do not want to compete with, that sentence needs rewriting. This step takes two minutes and catches most of what scanners miss.
3. Check your own site before publishing
Search your site for the draft's main terms. If an existing page covers the same question, the correct action is usually to improve and redirect to that page rather than publish a competitor to it. Two pages targeting the same query rarely both rank, and the newer one often cannibalises the older.
4. Verify every specific claim
For each number, date, name and quote: find the source and confirm it says what the draft says. Generated figures that look precise are the highest-risk content in any draft, because they read as researched and are not. Where a claim cannot be verified, cut it or replace it with something you can source.
5. Read the draft once, aloud, for provenance
Ask of each paragraph: where does this come from? If the answer is "the model", it needs a source or removal. This sounds slow, and it is — but in our testing it is the only step that reliably catches invented detail, and it needs doing once rather than for every page.
What a similarity score actually means
A percentage is a signal to investigate, not a verdict. Fourteen percent is normal for a page on a well-covered topic, because headings, terminology and standard definitions match by necessity. What matters is which passages matched and whether they carry the substance. Two percent on your core argument is worse than twenty percent on boilerplate.
Scores also move with length: short pages show higher percentages because any match is a larger share of the total. Compare like with like, and set thresholds by document type rather than using one number for everything.
How to fix what you find
Three responses, in order of preference. Attribute it — if the phrasing is worth keeping, quote it and cite the source. Consolidate it — if your own page already says it, strengthen that page. Rewrite it from your own material — test data, client examples, or specifics nobody else has. Replacing words with synonyms to get under a threshold produces text that is worse to read and still close in substance.
Where AI drafts need the most checking
In thirty days of testing, three content types produced the most overlap: definitional sections, where every page says roughly the same thing; tool feature lists, which copy vendor phrasing; and statistical claims, which are the most likely to be invented. Check all three manually even when the scan is clean.
Draft fast, verify slowly
Generation is the quick part. Budget more time for verification than for drafting — that ratio is what separates useful pages from filler.
Try Rytr →Frequently Asked Questions
What is an acceptable similarity score?
There is no universal threshold. Investigate each match instead: overlap in standard terminology is fine, overlap in your core argument is not.
Do AI detectors reliably identify AI content?
No. They produce false positives on human writing and disagree with each other. Do not use them as evidence.
Can AI output be plagiarised if it never copied directly?
Yes — near-duplicate phrasing and unattributed close paraphrase are the common cases, and scanners catch them poorly. The quoted-phrase search test works better.
How do I avoid duplicating my own pages?
Search your site for the draft's main terms before publishing. If a page already covers it, improve that page instead.
What should I do about unsourced statistics?
Verify or remove. A precise-looking figure that cannot be traced is the highest-risk sentence in a draft.
Is rewriting with synonyms enough?
Usually not. It lowers the score while keeping the substance close, and produces worse prose. Use attribution or your own material instead.