Cheap AI content is easy to spot.
Find out how yours reads.

It gives two separate readings and never blends them. How it reads is a deterministic measurement of the writing - sentence rhythm, point of view, vocabulary, stock phrases. Who wrote it is a trained model that makes one of three calls: written by AI, written by a person, or not called. It says AI only when a passage is unmistakably machine-written; when nothing is, it says so instead of guessing. A person can write generic copy and a good AI can write with voice - both readings can be true at once.

The text is cut into passages of about 300 words and each one is read by a classifier we trained on our own AI output and on writing we know people wrote. The document is judged by its most machine-like passage, because long AI drafts often have a machine-written opening and a body that reads human. The bar for an AI call is set high on purpose - on thousands of documents we know people wrote, almost none reached it - and there is a second model behind it that reports how predictable the writing is and which sentences pull the reading either way. Treat "not called" as exactly that: inconclusive, not a soft accusation. Paraphrased AI can still land there.

Higher is better writing. Seven measurements of the prose itself are weighted into one number: sentence rhythm counts most at 22 percent, stock phrases 20, point of view 14, short-sentence mix and active voice 12 each, facts versus interpretation and vocabulary range 10 each. Open the breakdown in any result to see all seven. The score says nothing about who wrote the text - that is the other reading.

The style score is: no model runs in it, and the same text always returns the same number, which is what makes a re-check after an edit worth anything. The who-wrote-it call comes from a model and is stable for the same text, but it is a call, not a measurement - which is why it is shown as words, never as a percentage.

It samples a domain's pages from its sitemap, reads each one the same way as a single check, and reports the share called AI, called a person, and not called - with a sampling interval, because thirty pages are a sample of the site, not the site. The per-page table shows which pages made the call.

At least 150 words for a reading you can rely on. Below that the ratios swing on a single sentence, and the result says so rather than pretending otherwise. The who-wrote-it call needs at least one full passage of about 300 words to be worth much; on a short text it will more often say "not called".

English only. Every word list and pattern in the style engine is English, and the origin model was trained on English, so a text in another language gets flagged instead of scored.

No. The text is sent to the scoring endpoint and to the model that reads it, measured, and never written anywhere. We keep only counts - how many checks, what the calls were, how long they took - never the text.