AI style analyzer without a score

The output is a written analysis of the submitted text: the register it occupies, how its sentences are constructed, which structures recur, where hedging displaces assertion, and where paragraph rhythm flattens. It is prose addressed to an editor, not a panel of gauges.

The route is server-side. Text is transmitted to our AI server for analysis and the result is returned. Input is accepted between 1 and 12,000 characters, the allowance is 5 free runs per rolling 24-hour window, and no account is required.

The established tools in this result set are metric engines. One computes word, sentence and character counts, a full punctuation frequency table, word and sentence length reported as mean, median and standard deviation, five named readability indices including Gunning Fog and Coleman-Liau, two readability graphs, lexical density with a parts-of-speech breakdown, and passive voice detection. Another adds tone classification against a fixed vocabulary, a sentiment figure and a suggested audience level.

That model has a real advantage: a number is comparable across drafts and across writers. It also has a persistent weakness, which is that a Gunning Fog index of 14 does not tell an editor which sentence to change. This route takes the opposite trade. It names passages and describes what they do, at the cost of producing nothing you can chart. Neither approach is strictly better, and a page selling one while implying the other is misleading its reader.

Submit a sample of several hundred words. A single sentence contains no distribution to observe.

Step 1

Paste the sample, from an essay, an article, a report or a page of documentation.

Step 2

Optionally specify an audience or constraint in the instruction field, then execute. The text is transmitted and analysed.

Step 3

Reconcile the report against the draft and apply what you accept. Remaining runs are shown on the interface.

Stated in advance so the report can be judged against a known scope rather than an assumed one.

Register and its stability

Which register the text occupies and whether it holds across sections or drifts between formal and promotional.

Sentence architecture

Verb position, clause stacking, subordination depth, and whether sentence length varies or converges.

Structural repetition

Recurring openers, connectives and paragraph shapes. Generally the highest-value finding in the report.

Hedging and assertion

Constructions that announce information rather than delivering it, and qualifiers that reduce a claim to nothing falsifiable.

Paragraph rhythm

Whether paragraphs develop an argument or enumerate. Convergence here is what readers describe as flat.

No readability index is computed: not Flesch-Kincaid, Gunning Fog, SMOG, Coleman-Liau or Automated Readability, and no Fry or Raygor graph is produced. No lexical density figure and no parts-of-speech tagging exist in this path. No punctuation frequency table, no word or sentence length distribution with mean, median and standard deviation, no passive voice ratio, no sentiment score.

No comparison against a corpus of named authors is performed, because no such corpus exists here. And unlike at least one competing style analyzer that advertises unlimited browser-based use, this route is neither local nor unlimited: the text is transmitted and the allowance is 5 runs per rolling 24 hours. Every item in this paragraph is a capability a competing page claims, and each one is absent here.

Next-token selection favours the probable, and probable constructions are those that fit the widest range of contexts. Applied across a document, that preference narrows the opener set, narrows the connective set, and drives paragraphs toward a repeated claim, elaboration and restatement cycle.

The resulting text contains no errors, which is precisely why error-focused tools report nothing useful about it. The defect is distributional rather than local, and it is only visible when a reader holds several paragraphs in view at once.

No. Style patterns are observable signals, not evidence of authorship. Formulaic human prose produces the same distribution, and carefully edited model output does not. The classifier is instructed never to assert certainty about authorship, plagiarism or which model generated a text, and to describe signals and limitations instead.

We make no prediction about any third-party detector's verdict, do not tune output against detector scoring, and do not test for it. The report is an editing instrument and nothing beyond that.

Editors triaging a draft they did not write, where the operative question is not correctness but why the piece reads thin. Content leads auditing a new contributor against an established house voice.

Students reviewing their own work ahead of submission, governed entirely by their institution's policy on AI-assisted tools. Anyone sufficiently close to a draft that they can no longer register its patterns unaided.