All benchmarks

How accurate is the plagiarism checker?

On 8 November 2026 we ran 56 test texts through the live checker, once on the Free plan and once on Go. The papers were drawn at random, and every result is here, including what it missed.

Read these results correctly

This is a small test: 30 copied passages, 8 light edits, 6 AI paraphrases, 4 short copies inside original writing and 8 original texts. With samples this size, one more found passage moves a figure by several points. It measures whether the report points at the right source. It does not compare our percentage with any other checker’s.

22of 30
Go plan · 19 of 30 on Free
Copied passages found

A 150-word passage copied word for word from a paper drawn at random. Found means the report named that paper or a copy of the same text elsewhere. 2 more were flagged through a different paper that contains the same text.

8of 8
Both plans · 0% and no sources
Original texts reported clean

Paragraphs written for this test on eight research topics. None was matched to anything, on either plan.

5of 8
Go plan · 3 of 8 on Free
Lightly edited copies found

Copied passages with about one word in eight swapped for a synonym, the way a student might edit before pasting.

Copied passages, by where they came from

SourceFound on GoFound on FreeFlagged through a related paper
Journal articles, 2025–268 of 106 of 101
Journal articles, 2010–245 of 65 of 6–
Theses3 of 63 of 61
Preprints3 of 42 of 4–
Wikipedia3 of 43 of 4–

A check of 150 words took 17 seconds at the median. Go searches the web for every sentence; Free searches a sample of them. Both read the same corpus of open research, so Go’s extra finds here (three copied passages and two light edits) came from the web search.

What it misses

Text reworded by an AI tool: 0 of 6

The checker compares exact wording. A full paraphrase keeps the meaning and changes the words, and this report does not detect it, on either plan.

Two copied sentences inside original writing: 0 of 4

A short copy is found only when those exact sentences are in our corpus or turn up in a web search. For these four they were in neither. Three of their papers were not found as whole passages either, and for the fourth the search engine returned no page for the two copied sentences: search engines do not index every sentence of every PDF. Given that thesis’s own text, the checker matched both sentences, so the gap is in finding the source.

Copied passages: 6 missed on Go

Two theses, two journal articles, a preprint from August 2026 and a Wikipedia article. Each is listed in the results file with its link.

Theses that repositories keep from automated readers

Many university repositories turn automated readers away. Of the 33 theses we drew at random, only 6 could be opened by any automated reader, so the test could only use those. A thesis its repository keeps from readers is invisible to this checker, and the theses figure above does not cover them.

How the test was built

The papers

Open-access works drawn at random from OpenAlex (a fixed random seed, so the draw can be repeated): journal articles from 2025–26 and from 2010–24, theses from 2018–26 and preprints from 2023–26, all labelled English, plus four random Wikipedia articles of at least 1,500 words. We skipped any work no automated reader could open or that held no 150 words of running prose: 6 of 16 journal articles from 2025–26, 4 of 10 journal articles from 2010–24, 27 of 33 theses, 7 of 11 preprints. The results therefore describe papers a reader can open. Three works labelled English were in Indonesian; we kept them.

The passages

From each work, the first 150 words of running prose from 40% of the way in (30% for Wikipedia), skipping tables, equations and reference lists. The rule was fixed before any check ran. Which passages were edited, paraphrased or mixed into original writing was decided by a seeded shuffle, not by hand. The light edits, paraphrases and original paragraphs were written by an AI model (Claude), as a student’s writing tool would write them.

The run

Each text went through the live checker at labrix.app as a signed-in user, one at a time, on the Free plan and on Go. The test accounts are left out of our public words-checked figure.

What counts as found

A source in the report is the paper itself (same DOI, OpenAlex record, arXiv number, address or title) or a copy of the same text at another address, such as a repository copy or a ResearchGate upload. We read every other match by hand. Two were a different paper that really contains the copied text, such as the author’s published version of a thesis chapter: we list those as flagged through a related paper, not as found. Two were short stock phrases from unrelated papers: we count those as missed.

Check it yourself

The results file lists every text: its type, the source paper with its link, and on each plan the outcome, the similarity score, the time taken and the top source the report named. It leaves out the passages themselves, because they quote other people’s papers; the links and the rule above let you rebuild them.

56 texts · 8,511 words · 8 November 2026.