Voice & tone

Standard deviation of sentence length in words, divided by the mean (the coefficient of variation). Reported instead of the mean because a mean of 18 describes both copy where every sentence is 18 words and copy that runs 4, 31, 7, 30 — and those read nothing alike.

Limits Scale-free by design, so it can be compared between sites. It says nothing about whether the long sentences are any good.

Mean absolute difference between each sentence and the one before it. Variation says the lengths differ; this says they interleave rather than arriving in one long block and then a short one.

Sentences are split on terminal punctuation followed by a capital, re-joining known abbreviations, then counted in words.

Limits A sentence with an unlisted abbreviation ("approx.", a middle initial) splits in two and counts as two short sentences.

Tokens are walked until the running type-token ratio drops to 0.72, which closes one factor; the count is tokens divided by factors, averaged forwards and backwards. Roughly: 50 is plain, 100 ordinary, 130 varied.

Limits Not scored below 50 tokens — there is not enough text for the number to mean anything. Higher is not automatically better: a site is allowed to call the same thing the same word every time, and should.

A direct count of each group’s pronoun list against the tokenized prose, reported per 10,000 words so documents of different lengths compare.

Limits "We" is counted wherever it appears, including inside a client quotation. A tagger would not help; only a human reading the sentence would.

A ratio of the two counts above. Null rather than infinity when the copy never addresses the reader, because "infinitely self-regarding" is a joke and not a measurement.

Limits On a case study the client is legitimately the subject, so a high ratio is evidence to read rather than a fault to fix.

A count of the booster list per 10,000 words. Appraisal calls this graduation: raising the force of a claim without adding anything a reader can check.

Limits A fixed list, so it undercounts. It also cannot tell a quoted booster from an asserted one.

A count of the hedge list per 10,000 words. Reported and not targeted: a little hedging reads honest, a lot reads evasive, and where the line falls is a house decision rather than a fact.

Modal verbs counted in three strengths. High closes the question, low holds it open, and copy with almost none of either is making bald assertions, which is a stance too.

Terminal punctuation decides questions and exclamations; a sentence opening with a bare verb from a short list counts as an imperative; everything else is declarative.

Limits The imperative list is deliberately short and excludes verbs that are also nouns, so "Design without direction is decoration" is not read as a command. It therefore undercounts.

Words ending -tion, -sion, -ment, -ance, -ence or -ism, per 10,000 words, minus a named list of lexicalised exceptions. Each one loses the person who did the thing: "we named it" against "the naming was undertaken".

Limits The exceptions list is by name, so an unlisted ordinary noun ("regiment", "torrent") would count as corporate prose. Add it to the list rather than living with the false positive.

A form of "be" or "get", optionally an adverb, then a participle — either regular (-ed, -en) or one from a list of irregulars.

Limits A heuristic, not a parse. It misses a prefixed irregular participle ("was redrawn", because the list holds "drawn"), and counts "is interested" as passive. Read it as an estimate. It is also not a fault to minimise: the passive is correct whenever the actor is unknown or beside the point.

A sentence counts as relational if it contains any copular or possessive verb. High density reads static: a list of what things are, with nothing happening. It is the grammar of an "about us" page.

Limits Detected by verb list, so a sentence with both a copula and an action verb counts as relational. The four process types overlap by construction and do not sum to the sentence count.

Any sentence containing a digit. Deliberately crude: the question is whether the copy counts anything at all, and a page that never does is a description rather than a result.

Limits A year counts, and a year is not an outcome. Read the share as a floor on how much evidence is present, never as a ceiling.

Contracted verb forms per 10,000 words. Bare ’s is ambiguous, so the hosts that genuinely contract a verb are named and possessives ("Miller’s bottle") are excluded.

First token of each sentence, matched against the coordinating conjunctions. The schoolroom rule against them is not an English rule; the number shows whether the site made a decision or never noticed.

Log-likelihood G² = 2·O·ln(O/E), where E is the word’s general-English rate times the corpus size. Filtered to words at least four times commoner than English (log-ratio ≥ 2), appearing four times or more, in at least three documents.

Limits Words absent from the reference list are dropped, not scored. They are nearly always brand names, and a table led by the client’s own name has answered its own question. The reference is a general-English blend, so a specialist site will look distinctive simply for being about its subject.

Unigram frequencies for the 17,626 commonest English words, in occurrences per million, from the wordfreq model — subtitles, news, books, web text and Wikipedia blended.

Limits General English, not a genre-matched corpus. Comparing marketing copy against a genre-matched reference would be sharper and needs a corpus the kit does not ship.

Words divided by strings, across every readable string in the corpus. Measured apart from prose rather than averaged into it: a caption is a different register from a paragraph, and scoring one for sentence rhythm reports a fault where there is a convention.

Strings are lowercased and counted; every string belonging to a group of two or more counts as repeated. A repeated caption is a caption that was not written for its image.

Limits Exact matches only, so two captions differing by one word both read as unique. Some repetition is honest — the same product shot appearing in four studies — so read a high share as a prompt to look, not as a fault.

A string counts as verbless if it contains no form of be or have, no modal, and no word ending -ed or -s. Reported and not targeted: a caption is allowed to be a noun phrase, and a site that captions in full sentences is making a different, equally valid choice.

Limits A regex, not a parse. A caption whose only verb is irregular ("Bottles made in Portland") reads as verbless, and a plural noun ("Bottles") reads as a verb.

Each of the catalogue’s ban-tier patterns matched case-insensitively across prose and display copy. Counted flat: one hit is a hit, and the table shows what it matched so the reader can judge it rather than take the count on trust.

Limits A hit is not evidence of anything. People write these too, and a handful across a whole site is a handful. The examples are shown for exactly this reason.

A density budget, not a ban: at most one hit per 400 words. Reported as words-per-hit against that budget. Position matters too — the catalogue counts a hit in a headline, an H1 or a CTA as over budget regardless of document length, so display-copy hits are counted separately.

Limits Chasing these to zero is the wrong outcome. Every one is a sentence a human writes sometimes, and a writer avoiding them all produces stilted prose, which is its own tell. Over budget means look, not delete.

Each shape matched by its own pattern over the prose, reported per 10,000 words against a budget. Shapes are measured because they survive paraphrase: a writer told to stop saying "not just X, it’s Y" writes "this isn’t merely X — it is Y", which a phrase list misses and the shape catches.

Limits The rule-of-three pattern is narrowed to three lowercase single words closing a sentence, because matching every "X, Y and Z" caught ordinary lists of services. It undercounts on purpose — for a number that gets read as an accusation, that is the right direction to be wrong in.

Every match is attributed to the field it came from as it is found, and display-copy matches are totalled across both tiers. Position is part of the measure rather than a refinement of it: the catalogue counts a hit in a headline, an H1 or a CTA as over budget regardless of how long the document is.

Limits It counts where a phrase sits, not how prominent that position actually is on the rendered page. A caption on the twelfth image and an H1 both count as display copy.

Each phrase matched case-insensitively on whole words, across prose and display copy together. A trailing YAML comment is stripped, and an entry written as a pattern ("it’s not just X, it’s Y") is skipped rather than matched literally.

Limits The site’s voice config is excluded from the corpus, or the avoid list would match itself once per entry. Only literal phrases are checked: a banned idea rephrased passes.