What we measured ourselves

Original ScriptGrain studies: what we measured ourselves, with the question, sample, method, findings and limitations stated in full.

• SGR-001: Same Brief, Four Voices
  Does a measured voice profile materially change the output of the same underlying model?
  Three content briefs, each run twice: once through a bare frontier model and once through the same model with a ScriptGrain voice profile applied. The facts stayed identical. The measured style did not.
• SGR-002: ChatGPT vs a measured voice
  When GPT-5 writes from the same brief three ways (bare, with a one-page measured prompt, and through ScriptGrain's profile), how close does each get to the author's own writing?
  One author, five published posts. Two were held out; a study profile was built from the other three. Each held-out post became a brief, and every brief ran through GPT-5 bare, GPT-5 with the prompt builder's system prompt, and ScriptGrain's generation. Every draft was measured with the product's own instruments against the study profile and against the held-out post it was meant to resemble.
• SGR-003: How many samples does a voice profile need before it stabilises?
  As pieces are added to a voice profile, when do its attributes stop moving, and when does it start telling its author from a stranger?
  Three public-domain authors (Dickens, Austen, Emerson), each cut into 800-word pieces. For each author, nested profiles were extracted from 1, 2, 3, 5, 8 and 15 pieces through the public API, then scored with the product's own voice-match instrument against four held-out pieces by the same author and eight by the other two. 18 profiles, 216 scores.
• SGR-004: Did ChatGPT change how companies write? 27 years of homepages, measured
  How has the writing on company homepages changed over 27 years, and did anything shift after ChatGPT arrived in November 2022?
  161 long-lived company homepages, one Wayback Machine capture a year from 2000 to 2026, rendered as a visitor saw them and measured with the same instruments the product uses. 2,950 homepages measured across 158 sites. The text was measured and discarded; only the numbers are published, as a free dataset under CC BY 4.0.
• SGR-005: Does a writer's voice change with age? 35 authors, 464 books, measured
  As writers age, which habits of their prose change and which stay fixed? Is there a general direction of change, such as longer sentences later in a career?
  464 books by 35 English-language authors, from 1810 to 1962, dated by first publication and measured passage by passage with the product's own instruments. Pre-registered. Two hypotheses tested, neither supported; published as a free dataset under CC BY 4.0.
• SGR-006: What does normal writing look like? Three centuries of human prose, measured
  For each habit of prose that can be counted, what is the normal range in human writing, by genre, and how has it moved from the 1700s to the web of 2021, before ChatGPT?
  A baseline for human writing: 5,720 texts and 6.0 million words, from 840 public-domain books in ten genres across three centuries to 2,200 pages of the 2021 web and 1,000 scientific abstracts, measured with the product's own instruments. Percentile tables for sixteen habits, three pre-registered hypotheses, and a free dataset under CC BY 4.0.
• SGR-007: Where did the AI-isms come from? 17,000 pieces of writing from before ChatGPT, measured
  Before ChatGPT, which kinds of human writing already used the words and habits now called AI-isms most, and when did each become common?
  17,303 pieces of writing and 10.1 million words from before ChatGPT, in 26 kinds of writing: the 2021 web that language models learned from (news in six countries, press releases, essay mills, marketing blogs, Wikipedia and more), 10,000 scientific abstracts, public-domain books by genre, and Google Books back to 1800. Pre-registered, measured with the product's own instruments, and published as a free dataset under CC BY 4.0.
• SGR-008: The AI tells moved: 2.1 million arXiv abstracts, 2015 to 2026
  When words like "delve" became a joke, did language models stop shaping scientific writing, or did the tells change?
  Every arXiv abstract from January 2015 to September 2026 (2,122,337 papers), measured against word lists fixed in advance. The 2024 AI tells peaked and fell below their pre-ChatGPT level; a different set of words rose in every field, so lists built on 2024 models now miss most of the signal.
• SGR-009: Do brands in the same sector sound the same? 89 UK brands' websites, measured
  Measured with a pairwise voice score, are two pages from the same brand's website more alike than a page from that brand and a page from a rival in the same sector?
  A pre-registered study of 299 pages from 89 UK brands' websites in 10 sectors, from banks to charities: homepage, about, product, pricing and careers pages, every pair scored with ScriptGrain's pairwise voice score and judged by two AI models in five runs. A brand's own pages were barely more alike than a rival's. Free dataset under CC BY 4.0.
• SGR-010: How do UK brands write? 16,070 blog posts and press releases, measured
  What is normal in the long-form writing of UK brands, by sector and by channel, and does a brand write the same way in its press releases as on its blog?
  A baseline for brand writing: 16,070 blog posts, press releases, announcements and reports from 117 UK brands in 10 sectors, 2015 to 2026, measured with ScriptGrain's instruments. Typical ranges for sentence length, "you", contractions, dashes and AI-associated words by sector and channel, two pre-registered comparisons, and a free dataset under CC BY 4.0.
• SGR-011: Did brand writing change after ChatGPT? 37 UK brands, before and after
  Within the same brands, did blog posts and press releases written after ChatGPT use more of the words and habits associated with AI writing than those written before it?
  A pre-registered before-and-after study of UK brands' own writing: each brand's blog posts and press releases from 2019 to 2022, read as published from the Wayback Machine, against the same brand's writing from 2024 to 2026. AI-associated words rose 40%; em dashes, AI sentence shapes and sentence length did not change. Free dataset under CC BY 4.0.
• SGR-012: Does a brand keep one voice across its press releases and blog? 14 UK brands, measured
  Measured with a writing-voice score, are two pages from the same brand and channel more alike than the same brand's blog and press releases, and is a rival's page in the same channel as close as the brand's own other channel?
  A pre-registered test on 1,074 blog posts and press releases from 14 UK brands, 2024 to 2026: 211,262 page pairs scored with ScriptGrain's counted-features voice score. A brand's pages in one channel agreed far more than its pages across channels, and a rival's page in the same channel came about as close as the brand's own other channel. Free data under CC BY 4.0.
• SGR-013: How do UK banks and building societies write? 18 brands measured against nine other sectors
  How does the writing of UK banks and building societies differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?
  2,669 blog posts and press releases from 18 UK banks and building societies, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.
• SGR-014: How do UK energy suppliers and energy brands write? 11 brands measured against nine other sectors
  How does the writing of UK energy suppliers and energy brands differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?
  1,949 blog posts and press releases from 11 UK energy suppliers and energy brands, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.
• SGR-015: How do UK mobile, broadband and TV providers write? 15 brands measured against nine other sectors
  How does the writing of UK mobile, broadband and TV providers differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?
  1,816 blog posts and press releases from 15 UK mobile, broadband and TV providers, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.
• SGR-016: How do software companies with UK sites write? 16 brands measured against nine other sectors
  How does the writing of software companies with UK sites differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?
  3,422 blog posts and press releases from 16 software companies with UK sites, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.
• SGR-017: How do UK insurers and insurance brands write? 10 brands measured against nine other sectors
  How does the writing of UK insurers and insurance brands differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?
  2,049 blog posts and press releases from 10 UK insurers and insurance brands, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.
• SGR-018: How do UK charities write? 10 brands measured against nine other sectors
  How does the writing of UK charities differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?
  1,994 blog posts and press releases from 10 UK charities, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.
• SGR-019: Has the em dash really taken over? Em dashes per 1,000 words in books, research papers and brand writing, 1800 to 2026
  How often did published writing use the em dash from 1800 to 2026, did its use in brand writing, research abstracts and company homepages rise after ChatGPT beyond its earlier trend, and is today's rate high or low against 19th-century books?
  A pre-registered count of em dashes per 1,000 words across 2,059,475 arXiv abstracts, 7,464 dated brand pages, 2,950 company homepages and 1,276 Project Gutenberg books. Brand writing: 2.72 in 2019 to November 2022, 2.28 since 2024. Nineteenth-century non-fiction books: 4.89.
• SGR-020: Did government writing change after ChatGPT? 55,350 GOV.UK press releases and news stories, 2015 to 2026
  Within the same UK government departments, did press releases and news stories published after ChatGPT use more of the words and habits associated with AI writing than those published before it?
  A pre-registered before-and-after study of 55,350 GOV.UK press releases and news stories from 2015 to September 2026. Each of 20 departments is compared with itself: 2019 to November 2022 against 2024 onwards. AI-associated words rose 22%; em dashes did not. Free dataset under CC BY 4.0, source credited under the Open Government Licence.
• SGR-021: Which classic authors write most like an AI? 1,129 authors, 6,293 books, measured
  Which human writers, long before any language model, already used the words and habits now called AI-isms, and how much? Which of those words lived in which authors and eras?
  6,293 books by 1,129 English-language authors from Project Gutenberg, ranked by how often they use the word lists now called AI-isms, with intervals over their books and a check on era, genre and dialogue. Pre-registered; published as a free dataset under CC BY 4.0. It ranks human habits; it does not detect AI.
• SGR-022: Do ministers' written answers sound different since ChatGPT? 42,082 parliamentary answers and 17,292 written statements, 2015 to 2026
  Within the same UK government departments, do written parliamentary answers and written ministerial statements given after ChatGPT use more of the words and habits associated with AI writing than those given before it?
  A pre-registered before-and-after study of 42,082 written answers to parliamentary questions and 17,292 written ministerial statements from 2015 to September 2026, drafted largely by civil servants. Each government department is compared with itself: 2019 to November 2022 against 2024 onwards. AI-associated words rose 26% in answers; em dashes did not. Free dataset under CC BY 4.0; source credited under the Open Parliament Licence.
• SGR-023: What reading age do UK brands write for? Readability of 115 brands' blogs and press releases, 2015 to 2026
  What Flesch-Kincaid grade and UK reading age do UK brands' blog posts and press releases demand, how do sectors differ, are press releases harder to read than the same brand's blog posts, and did readability change between 2015 and 2026, including after the FCA's Consumer Duty came into force?
  A pre-registered readability measurement of 15,540 blog posts, press releases and announcements from 115 UK brands in 10 sectors. The median brand's writing scores Flesch-Kincaid grade 10.7, a reading age of 15.7; supermarkets are hardest (18.4) and food and drink easiest (14.5).

SGR-001: Same Brief, Four Voices

Does a measured voice profile materially change the output of the same underlying model?

Three content briefs, each run twice: once through a bare frontier model and once through the same model with a ScriptGrain voice profile applied. The facts stayed identical. The measured style did not.

Conducted 2026-07-27, published 2026-07-28.

Sample

Three original content briefs (a banking feature announcement, an opinion piece on productivity, and a B2B piece on replacing status meetings), each run through two arms. A fourth run reused the banking brief against a profile built from a Victorian novelist's prose. (n = 8)

Method

• Arm one: the same frontier model ScriptGrain runs on, called bare, with no system prompt and nothing but the brief.
• Arm two: the ScriptGrain pipeline with a voice profile applied as a hard constraint on generation.
• Outputs were compared on measured style features rather than on preference, so the comparison does not rest on which version anyone liked.

What was measured

• Average sentence length
• Share of sentences under eight words
• Contraction rate
• Em-dash count
• Formatting shape (headings, bolded topic sentences, list use)
• Output length against the brief's word target

Findings

• Sentence length collapsed on the challenger-bank brief: 17 to 11 words. Average sentence length in the profiled arm fell to roughly 11 words from roughly 17 in the bare arm, and sentences under eight words doubled from 24% to 48%.
• Voice lives in the distribution, not the average: 16% to 36%. On the opinion brief the average sentence length was effectively identical across arms at about 15 words, yet short punch sentences went from 16% to 36%. Averaging the two arms together would have shown no difference at all.
• Em dashes vanished: 10 to 0. The bare arm used ten em dashes on the banking brief. The profiled arm used none, which is the single most visible marker separating default model prose from written-by-a-person prose.
• Formatting is part of identity: 708 to 605 words. The bare arm produced a scannable listicle with four headings and fourteen bolded topic sentences, and overran a 600-word brief by 108 words. The profiled arm produced continuous prose with no headings and landed at 605.
• An extreme profile holds under the same facts: 0 contractions in 626 words. The banking brief run against a Victorian-novelist profile carried the identical feature, mechanic and price point, with no contractions, no headings and sums of money spelled out as words.

Limitations

• One run per arm. This is a demonstration of effect direction and size on a handful of briefs, not a statistical result, and no significance test was performed.
• The briefs were written by the study's author, who also builds the product. Neither arm was blind.
• Both arms used the same underlying model, so the study says nothing about how profiles behave across different models.
• A voice profile changes identity, not quality. Nothing here shows that profiled output is better, only that it is measurably different and consistently so.

Read the full study

SGR-002: ChatGPT vs a measured voice

When GPT-5 writes from the same brief three ways (bare, with a one-page measured prompt, and through ScriptGrain's profile), how close does each get to the author's own writing?

One author, five published posts. Two were held out; a study profile was built from the other three. Each held-out post became a brief, and every brief ran through GPT-5 bare, GPT-5 with the prompt builder's system prompt, and ScriptGrain's generation. Every draft was measured with the product's own instruments against the study profile and against the held-out post it was meant to resemble.

Conducted 2026-09-20, published 2026-09-20.

Sample

Five posts (5,233 words) from one author's company blog, already the samples of an existing ScriptGrain profile. The two longest (1,394 and 1,392 words) were held out; a fresh study profile was extracted from the remaining three (2,447 words). Two briefs, three arms, six drafts. (n = 6)

Method

• Study profile: api-v1 POST /v1/profiles from the three training posts (one extraction slot), status polled to complete.
• Briefs: gpt-5-mini read each held-out post and wrote the brief a writer would have received (topic, angle, audience, three points, target length), instructed not to quote it.
• Arm one, bare: GPT-5 (reasoning effort minimal) with the brief and a target length only.
• Arm two, prompt: GPT-5 with a system prompt from ScriptGrain's free prompt builder, measured from the three training posts (numbers for sentence length and variation, contractions, punctuation, pronoun mix; judged formality, humour, argument structure; the hard rules against AI tells).
• Arm three, ScriptGrain: api-v1 POST /v1/generations against the study profile, content type article, the brief's target length.
• Measures per draft: voice match against the study profile (POST /v1/voice-match, feature-delta engine), pairwise match against the held-out post (POST /v1/compare-voice), AI-tell density and em dashes (the public cliché checker's scanner), share of sentences under eight words, and the AI detector's human score (POST /v1/ai-detect). Runner: scripts/experiments/run-study.mjs in the ScriptGrain repository.

What was measured

• Voice match against the study profile (0 to 1, profile-calibrated)
• Pairwise match against the held-out post (0 to 1, pairwise-calibrated)
• AI-tell phrases and sentence shapes per 1,000 words
• Em dashes per draft
• Share of sentences under eight words
• Sentence-length variation (standard deviation over mean)
• Detector human score (0 to 1)
• Draft length against the brief's target

Findings

• The measured profile got closest to the author: 0.83 vs 0.74. Mean voice match against the study profile was 0.83 for ScriptGrain's drafts, 0.74 for bare GPT-5 and 0.62 for GPT-5 with the one-page prompt. Against the held-out post itself the order was the same: 0.91, 0.82, 0.80.
• A one-page prompt was not enough for GPT-5: 0.62. The prompt asked for about 14 contractions per 1,000 words and a formality of 3.7 on a 0 to 10 scale; GPT-5 produced 0.5 contractions per 1,000 and a judged formality of 7.2 on both briefs, more formal than its bare drafts. It did obey the hard rules (zero em dashes, more short sentences, higher sentence variation), so the prompt shaped the surface and missed the register.
• Bare GPT-5 ignored the length and reached for em dashes: 2309 words, 17.5 em dashes. Asked for about 1,393 words, bare GPT-5 wrote 2309 on average with 17.5 em dashes per draft. ScriptGrain's drafts averaged 1302 words with none; the prompt arm also removed every em dash.
• ScriptGrain's own drafts still produced 'not X but Y': 2.3 per 1,000. The checker counted more AI-tell constructions in ScriptGrain's drafts (mean 2.3 per 1,000 words; 3.8 on the first brief, three 'not X, not Y, but Z' sentences) than in bare GPT-5's (0.2). The generation prompt bans the construction and the model wrote it anyway.
• Detector scores did not track voice: 0.78 / 0.67 / 0.52. The AI detector read the prompt arm as most human (0.78), then ScriptGrain (0.67), then bare GPT-5 (0.52, with one draft at 0.32). Sounding human and sounding like the author are different measurements; the study reports both.

Limitations

• Two briefs, one author, one run per arm. Effect direction and rough size only; no significance test.
• The study profile was built from 2,447 words across three posts, below the 3,000 words the product recommends, so both the profile and the prompt were thinner than a real user's would be.
• No author-written custom-instructions arm; the author had not written any. It is the first addition for the next run.
• Briefs were derived from the held-out posts by a model, so they may carry more of the piece's structure than a human brief would.
• The author of the held-out posts also builds the product. Nothing was blind, and no draft was chosen or edited after the run.
• Pairwise scores were read on the scale in use when the study ran. On 29 September 2026 that scale was re-set from 558 measured pairs, because pairs by different writers had read "on voice" 40% of the time. Re-read on the new scale, the same runs average 0.84 for ScriptGrain (on voice), 0.70 for bare GPT-5 and 0.67 for GPT-5 with the prompt (both drifting). The order is unchanged.
• The AI-tell scanner is ScriptGrain's own instrument and the reference bands come from its reference corpus; another checker would count differently.

Read the full study

SGR-003: How many samples does a voice profile need before it stabilises?

As pieces are added to a voice profile, when do its attributes stop moving, and when does it start telling its author from a stranger?

Three public-domain authors (Dickens, Austen, Emerson), each cut into 800-word pieces. For each author, nested profiles were extracted from 1, 2, 3, 5, 8 and 15 pieces through the public API, then scored with the product's own voice-match instrument against four held-out pieces by the same author and eight by the other two. 18 profiles, 216 scores.

Conducted 2026-09-20, published 2026-09-21.

Sample

Great Expectations (Gutenberg #1400), Pride and Prejudice (#1342) and Emerson's Essays (#16643), body text only, cut at paragraph boundaries into pieces of about 800 words (600 to 1,400 accepted). With a fixed seed (20260920), 15 training pieces and 4 held-out pieces were drawn per author; the training sets are nested, so the 1-piece profile's sample is inside the 2-piece profile's, and so on. Public-domain prose was used so the study is reproducible from the source text and needs nobody's consent. (n = 18)

Method

• Profiles: POST /v1/profiles with the first k training pieces (k = 1, 2, 3, 5, 8, 15), status polled to complete. Extraction is the same two-pass pipeline every user gets; extractions are rate-limited to five an hour, so the run spans several hours.
• Same-author score: POST /v1/voice-match for each of the author's four held-out pieces against the k-piece profile; the mean is reported.
• Other-author score: the same call for the other two authors' eight held-out pieces; the mean is reported. The gap (same minus other) is how well the profile tells its author from a stranger.
• Attribute agreement: each k-piece profile's enumerated attributes (complexity, clause ordering, humour, expressiveness, opening style, argument structure, ellipsis habit, list preference, connective preference, specificity, rhythm, paragraph length, power-sentence position) compared with the author's 15-piece profile; the share that agree is reported. Numeric attributes are reported as the absolute difference from the 15-piece value.
• Confidence: the confidence_score the extraction itself returns with each profile.
• Runner: scripts/experiments/run-stability-study.mjs in the ScriptGrain repository; data file linked below.

What was measured

• Voice match of held-out pieces by the same author (0 to 1, profile-calibrated)
• Voice match of held-out pieces by the other two authors (0 to 1)
• Discrimination gap (same minus other)
• Share of enumerated attributes agreeing with the 15-piece profile
• Absolute drift of numeric attributes from the 15-piece profile
• Extraction confidence score (0 to 1)
• Words in the training set

Findings

• The score barely cares how many pieces built the profile: 0.86 at 1 piece, 0.86 at 3, 0.86 at 15. Mean voice match of each author's four held-out pieces against their own profile, averaged over three authors, did not move as pieces were added: 0.86 at one piece, 0.86 at three, 0.86 at fifteen. A single 800-word piece already pins the counted habits (sentence length, contractions, commas, pronouns) closely enough for the scorer to read the author's other work as on voice.
• What more pieces buy is a profile that stops moving: sentence length 4.9 words off at 1 piece, 1.2 at 3; contractions 3.6 per 1,000 off, then 0.8. Measured against each author's fifteen-piece profile, the one-piece profile's average sentence length was off by 4.9 words, contractions by 3.6 per 1,000 and formality by 0.9 points on the 0 to 10 scale. At three pieces those errors were 1.2 words, 0.8 per 1,000 and 0.3; from five pieces on, under two words, under one contraction per 1,000 and about 0.1 on formality.
• The extraction's own confidence tracks sample size: 0.67 at 1 piece, 0.79 at 3, 0.87 at 15. The confidence_score returned with each profile rose with the pieces behind it: 0.67 on average at one piece, 0.78 at two, 0.79 at three, 0.80 at five, 0.86 at eight, 0.87 at fifteen.
• Telling one author from another is where the method is weakest here: gap 0.06 to 0.10; Dickens 0.17, Emerson 0.01. The gap between an author's own held-out pieces and the other two authors' averaged 0.06 to 0.10 at every sample count, with no trend. Dickens separated best (0.17 from five pieces), Emerson hardly at all (0.01 at fifteen: his profile scored Dickens and Austen as high as it scored him). All three are formal, dry, flowing nineteenth-century voices at formality 7 to 9 with almost no contractions, so the counted habits the score leans on are shared; the scorer measures distance on those habits and reads similar registers as similar. A modern conversational writer scored against any of these profiles would likely sit far lower, though this study did not test one.
• Labelled attributes are the least stable part of a profile: 72% to 87% agreement with the 15-piece profile. The share of thirteen enumerated attributes (humour register, argument structure, rhythm, clause ordering and the rest) that matched the fifteen-piece profile's labels ran 82% at one piece, 72% at two, 82% at three, 74% at five and 87% at eight, with Emerson as low as 46% at two pieces. These labels are a model's reading of the samples and flip between extractions more than the counted numbers move. Eight pieces or more is where they settle.

Limitations

• Three authors, one draw of pieces each, one extraction per profile: the direction of the curves is the finding, not the decimals. A second draw would move individual cells.
• All three authors are nineteenth-century literary prose, which the scorer separates less sharply than it separates a newsletter from a legal notice; the other-author baseline is therefore a hard test, and a modern business writer's profile would likely show a larger gap.
• The judged attributes come from a model reading the samples, so some run-to-run variation in labels is inherent and is part of what this study measures.
• Pieces of about 800 words are longer than a typical LinkedIn post and shorter than a typical blog post; the piece count is the variable studied, and the word count scales with it.
• The author of this study builds the product. Nothing was blind; nothing was re-run or selected.

Read the full study

SGR-004: Did ChatGPT change how companies write? 27 years of homepages, measured

How has the writing on company homepages changed over 27 years, and did anything shift after ChatGPT arrived in November 2022?

161 long-lived company homepages, one Wayback Machine capture a year from 2000 to 2026, rendered as a visitor saw them and measured with the same instruments the product uses. 2,950 homepages measured across 158 sites. The text was measured and discarded; only the numbers are published, as a free dataset under CC BY 4.0.

Conducted 2026-10-04, published 2026-10-04.

Sample

A fixed panel of 161 companies whose homepages have been online since 2000 or soon after, chosen before anything was measured: 11 sectors (software, hardware, retail, finance, insurance, telecoms and media, travel, consumer brands, logistics and cars, agencies and consultancies, web and marketing tools), UK and US. For each site and each year, the homepage was requested from the Wayback Machine at 1 July; the nearest capture was kept only if it fell in the same calendar year. Each capture was then rendered with JavaScript (so a homepage built in the browser is measured on what a visitor saw, not an empty shell), and the raw archived HTML was measured too, as a check. (n = 2950)

Method

• Panel and vocabularies fixed on 4 October 2026, before any page was measured (scripts/experiments/sgr-004/panel.json and lexicons.json).
• Fetch: the Wayback Machine's raw capture for each site-year (crawl.mjs), then the same capture rendered with JavaScript by Firecrawl with the archive's toolbar removed (render.mjs). Two thirds of the way through, the archive rate-limited our server; the remaining site-years were requested by date through Firecrawl only, so they have no raw-HTML twin.
• Extraction: visible prose lines, with scripts, navigation and footers removed, using the same block splitting and prose-line rule as the product's URL reader. Unlike that reader, page headers, forms and asides are kept, because on a homepage that is where the headline copy sits; repeated lines (desktop and mobile copies of one carousel) are counted once.
• Measurement: the product's own deterministic instruments, measureTextFeatures and scanCliches, the same code that powers the AI cliché checker and voice matching. No language model was used to read or score any page.
• Vocabularies: an AI-era list (29 entries, among them seamless, unlock, elevate, streamline) and a dotcom corporate list (19 entries, among them solutions, world-class, synergy, click here), counted per 1,000 words. Neither says a page was written by AI; they track vocabulary only.
• Before and after: each site's mean over its 2019 to 2022 captures against its mean over its 2023 to 2026 captures, paired by site, so sites entering or leaving the panel cannot move the result. ChatGPT launched on 30 November 2022; captures are taken nearest 1 July, so 2022 is before and 2023 after.

What was measured

• Words of prose on the homepage
• Average sentence length and its variance
• Contractions, exclamation marks and em dashes per 1,000 words
• Share of I, we and you among first and second person pronouns
• AI-era and dotcom vocabulary per 1,000 words
• ScriptGrain AI-tell density per 1,000 words (phrases and sentence shapes)

Findings

• AI-era marketing words nearly doubled after ChatGPT: 1.05 to 1.93 per 1,000 words; 83 sites up, 27 down. Pairing each site's own homepage before and after November 2022, the AI-era vocabulary (style words only) rose from 1.05 to 1.93 uses per 1,000 words on average: up on 83 of 147 sites, down on 27, unchanged on 37. "Unlock" went from 7.1 to 31.3 uses per 100,000 words, "seamless" from 8.0 to 16.6, "streamline" from 3.1 to 8.5. The rise started before ChatGPT: the yearly average was 0.45 in 2012 and 1.29 in 2019. After 2022 it reached 2.17 in 2026.
• The dotcom vocabulary faded, and the AI-era words overtook it in 2023: 3.65 per 1,000 words in 2000, 1.77 in 2026. "Solutions", "world-class", "click here", "welcome to" and the rest of the 2000s corporate vocabulary averaged 3.65 uses per 1,000 words in 2000 and 1.77 in 2026. Most of that fall came early: paired by site, the dotcom rate was 1.61 before ChatGPT and 1.60 after. So when the AI-era words became the more common of the two in 2023, and stayed so every year since, it was because they rose, not because the old words fell further. One vocabulary of borrowed confidence replaced another: in 2000 homepages said "solutions", in 2026 they say "seamless".
• AI did not bring the em dash to homepages: 3% of homepages in 2000, 27% in 2022, 38% in 2026. The share of homepages using at least one em dash climbed for two decades before ChatGPT: 3% in 2000, 16% in 2010, 27% in 2022. It kept rising after, to 38% in 2026: 2.7 points a year since 2022 against 1.1 a year before it. Paired by site, the em dash rate per 1,000 words went from 1.15 before ChatGPT to 1.24 after (up on 52 sites, down on 38). The habit was already well under way; AI may have quickened its spread, but it did not start it.
• The sentence shapes people call AI did not arrive on homepages: 0.16 before, 0.11 after, per 1,000 words. ScriptGrain's AI-tell density (stock transitions such as "furthermore" and "delve into", plus shapes such as "not X but Y" and the reflexive list of three) was low throughout and did not rise: 0.16 per 1,000 words before ChatGPT, 0.11 after, with 106 of 147 sites at the same level. Homepage copy is short and headline-shaped, which leaves little room for those shapes; the change shows in the words, not the sentences.
• Homepages got longer and a little less conversational: median 169 words in 2000, 606 in 2026. The median homepage carried 169 words of prose in 2000, 319 in 2010, 481 in 2022 and 606 in 2026. Paired by site, 101 of 147 homepages were longer after ChatGPT than before. Sentences lengthened slightly (median 7.9 words in 2000, 9.0 in 2026), and contractions fell after 2022 on 85 of 147 sites (11.9 to 11.5 per 1,000 words).

Limitations

• Homepages only. A homepage is the most edited, most headline-shaped page a company has; blog posts and product pages may have moved differently.
• The panel is 161 large, long-lived companies, mostly UK and US. Small businesses and companies founded after 2010 are not represented, and a young company's copy may look different.
• Vocabulary is not authorship. A rise in 'unlock' and 'seamless' is consistent with AI-assisted writing, and also with copywriters adopting the words that AI made familiar. This study cannot tell those apart, and it does not try to say which homepage was written by a model.
• 'AI-powered' and 'powered by AI' were on the AI-era list as fixed before measuring. They mostly describe what a company sells, not how its copy reads, so the headline figures exclude them. That exclusion was decided after the list was fixed and is disclosed here; the figure with them left in is in Checks and in the data.
• The Wayback Machine keeps what it captured, when it captured it. Some captures are partial, some years have none, and the nearest capture to 1 July can be weeks away from it. Rendering an old capture runs its archived scripts, and some of those no longer load.
• Medians and means over a changing set of sites can move because the set changed. The paired before and after comparison and the fixed-panel check exist for that reason; the yearly lines should be read as the broad trend.

Read the full study

SGR-005: Does a writer's voice change with age? 35 authors, 464 books, measured

As writers age, which habits of their prose change and which stay fixed? Is there a general direction of change, such as longer sentences later in a career?

464 books by 35 English-language authors, from 1810 to 1962, dated by first publication and measured passage by passage with the product's own instruments. Pre-registered. Two hypotheses tested, neither supported; published as a free dataset under CC BY 4.0.

Conducted 2026-10-04, published 2026-10-04.

Sample

Wikidata records (query and saved result in the repository) for English-language works with a Project Gutenberg ID, an author with birth and death dates, and a first-publication date. A date counts only if it falls between the author's 15th birthday and two years after their death; later dates are editions. Drama and poetry are excluded by Project Gutenberg subject, and every book has at least 15,000 words. 35 authors met the bar of at least six books spanning fifteen years or more: 464 books from 1810 to 1962, five passages each (2,335 passages, 2.5 million words). (n = 464)

Method

• Pre-registration: the hypotheses, panel rules and measures were committed to the public repository before any book was measured (docs/research/sgr-005-preregistration.md). The panel is whatever those rules produce, with no hand edits.
• Measurement: each passage scored with the same deterministic code ScriptGrain uses, once as narration (dialogue removed) and once as all text; a book's value is the mean of its five passages. No language model reads or scores any text.
• H1: for each author, a straight-line slope of narration sentence length against age; the mean of the slopes, with a 95% bootstrap interval over authors.
• H2: for six core habits, the intraclass correlation of book values within authors, with a 95% bootstrap interval over authors; supported if at least four of the six intervals sit above 0.5.

What was measured

• Average sentence length
• Sentence length variation
• Commas per sentence
• Semicolons per 1,000 words
• Contractions per 1,000 words
• Share of long words
• Exclamation marks and dashes per 1,000 words
• Pronoun shares
• Share of words in dialogue

Findings

• H1 (not supported): sentences did not get longer with age in general: −0.1 words per sentence per decade; 95% interval −0.7 to +0.5. 14 of 35 authors' narration sentences lengthened as they aged and 21 shortened. Counting dialogue too gave the same answer (+0.0 per decade, interval −0.5 to +0.7).
• Writers drifted in their own directions: Henry James +6.0, Elizabeth von Arnim −3.4 words per decade. Henry James's late style is real: his narration sentences averaged 16.9 words in 1875 to 1886 and 32.0 in 1901 to 1909. George Gissing (+4.2) and E. M. Forster (+2.0) also lengthened. Others went the other way: Elizabeth von Arnim (−3.4), Anthony Trollope (−2.6; without his first two novels, of 1847 and 1848, his trend is +1.0, a reminder that one writer's line can rest on a few books) and Charles Dickens (−1.8).
• H2 (not supported): habits were only partly fixed: 0.34 to 0.57 of the variation explained by the author. For a stable fingerprint we required the author to explain more than half of the variation, with confidence, in at least four of six core habits. None cleared that bar. Long words (0.57), semicolons (0.56), commas per sentence (0.56) came closest, with estimates above 0.5 but intervals reaching below it; sentence length (0.39), contractions (0.39), sentence length variation (0.34) varied more between an author's own books.
• Exploratory: late books mixed long and short sentences more: 23 of 35 authors; +1.5 points of variation per decade. Sentence length variation rose with age for 23 authors and fell for 12, a small change whose interval sits just above zero. Semicolons tended to fall (23 of 35 authors), though that interval includes zero. These were not pre-registered and should be read as leads.

Limitations

• Project Gutenberg texts come from particular editions, and punctuation can be an editor's or printer's rather than the author's; semicolons and dashes are the most exposed.
• Wikidata's first-publication dates are mostly right but not checked book by book; a wrong date moves one point on one author's line.
• 64 books use single quotation marks for dialogue, or have none; for those, the narration measure still contains dialogue.
• Authors write different kinds of books at different times (novels, stories, travel, essays), so some change across a career is a change of form, not of voice.
• Thirty-five authors, nearly all British or American and writing between 1810 and 1962; the findings describe them, not writers in general.

Read the full study

SGR-006: What does normal writing look like? Three centuries of human prose, measured

For each habit of prose that can be counted, what is the normal range in human writing, by genre, and how has it moved from the 1700s to the web of 2021, before ChatGPT?

A baseline for human writing: 5,720 texts and 6.0 million words, from 840 public-domain books in ten genres across three centuries to 2,200 pages of the 2021 web and 1,000 scientific abstracts, measured with the product's own instruments. Percentile tables for sixteen habits, three pre-registered hypotheses, and a free dataset under CC BY 4.0.

Conducted 2026-10-04, published 2026-10-04.

Sample

Books: Project Gutenberg English texts in ten genres (fiction, essays, history, science, philosophy and religion, travel, biography, letters, children's books, drama; poetry excluded), 30 per genre and era where at least 30 exist, in a seeded random order, with three passages of about 1,000 words at 20%, 50% and 80% of each book. Three cells (1700s essays, children's books and drama) had too few books and were dropped, not filled. Web: 100 pages from each of 22 sources in nine kinds of writing, from Common Crawl's October 2021 crawl, in SGR-007's frozen order. Abstracts: 1,000 arXiv abstracts from 2015 to November 2022 with no later revision, from SGR-008's census. (n = 5720)

Method

• Pre-registration: hypotheses, genres, eras and sampling were committed to the public repository before any text was measured (docs/research/sgr-006-preregistration.md).
• Measurement: every text scored with the same deterministic code ScriptGrain uses, for sixteen habits, on all text. No language model reads or scores any text.
• Re-run: the first books harvest was found to contain rows written by two runs at once, with some books mislabelled. It was discarded before analysis and the books were harvested again by a single run; the harvester now refuses to write into an existing file.
• Check added after the main run (not pre-registered): web pages and the non-fiction book passages re-measured on full prose lines only (8 words or more, ending in sentence punctuation), to see whether the web's shorter sentences come from headings and lists.

What was measured

• Words per sentence, sentence length variation, share of short sentences
• Commas per sentence; semicolons, exclamation marks, question marks and dashes per 1,000 words
• Contractions per 1,000 words; parentheticals
• Short-word and long-word shares
• I, we and you shares of pronouns
• AI-tell density

Findings

• H2 (supported): non-fiction sentences got shorter, century after century: 32.2 → 28.6 → 24.5 → 15.1 words per sentence. Early-1900s non-fiction books averaged 7.7 fewer words per sentence than 1700s books (95% interval 6.4 to 9.0 fewer), and 2021 web non-fiction 9.4 fewer again. Part of the web figure comes from headings and lists counted as sentences; on full prose lines only, news and how-to pages averaged 19.3 words against 25.0 for early-1900s books, a smaller gap in the same direction.
• H3 (not supported): contractions did not creep up in books; they jumped on the web: 4.4 → 4.1 in books, 13.9 on the 2021 web, per 1,000 words. We predicted a rise from the 1700s to the early 1900s. Non-fiction books were flat (difference −0.3, interval −1.0 to +0.4). The jump came after: 2021 web non-fiction used 9.7 more per 1,000 words than early-1900s books, and still 15.4 per 1,000 on full prose lines only.
• H1 (not supported): fiction was not always the short-sentence genre: 1700s fiction 36.6 words per sentence; 1800s 21.3; early 1900s 17.6. In the 1800s and early 1900s, fiction's sentences were shorter than history, science and philosophy, as predicted. In the 1700s they were not: novels ran 5.8 words longer than history, and the comparisons with science and philosophy and religion were inconclusive. Fiction's sentences shortened faster than any non-fiction genre's, by about 19 words between the 1700s and the early 1900s.
• Exploratory: semicolons faded and dashes rose, then fell: semicolons 12.1 → 1.5; dashes peaked at 6.6 in the early 1900s. In non-fiction, semicolons fell in every era, from 12.1 per 1,000 words in the 1700s to 5.9 in the early 1900s and 1.5 on the 2021 web. Dashes rose from 2.7 to 6.6 and were rare on the 2021 web (1.3): books of a century ago used them about five times as often as web pages did just before ChatGPT. Not pre-registered; read as description.

Limitations

• Books are dated by when their author turned 40, not by publication; a book written young or old can sit in a neighbouring era.
• Project Gutenberg texts come from particular editions, and punctuation can be an editor's or printer's; semicolons and dashes are the most exposed.
• The web's sentence measure counts headings, captions, list items and references as short sentences, which pulls its figures down, most for Wikipedia (see Checks).
• Genres are assigned from catalogue subjects, so a book can sit in more than one (travel letters, for example) and some labels are broad.
• English only, and mostly British and American writing.

Read the full study

SGR-007: Where did the AI-isms come from? 17,000 pieces of writing from before ChatGPT, measured

Before ChatGPT, which kinds of human writing already used the words and habits now called AI-isms most, and when did each become common?

17,303 pieces of writing and 10.1 million words from before ChatGPT, in 26 kinds of writing: the 2021 web that language models learned from (news in six countries, press releases, essay mills, marketing blogs, Wikipedia and more), 10,000 scientific abstracts, public-domain books by genre, and Google Books back to 1800. Pre-registered, measured with the product's own instruments, and published as a free dataset under CC BY 4.0.

Conducted 2026-10-04, published 2026-10-04.

Sample

33 web sources in 16 kinds of writing, read from Common Crawl's October 2021 crawl (CC-MAIN-2021-43): up to 200 pages per source, in a seeded random order, kept if the product's article reader found at least 200 words. The 103 company homepages are the SGR-004 panel as Common Crawl saw them in October 2021. Scientific abstracts: 10,000 arXiv abstracts, a seeded random sample of the 1,057,471 from 2015 to November 2022 with no revision after 30 November 2022 (SGR-008's census). Books: 25 public-domain books in each of 8 genres from Project Gutenberg, three passages of about 800 words from each. Google Books Ngram (English 2019 corpus) for every listed word, 1800 to 2019. (n = 17303)

Method

• Pre-registration: hypotheses H1 to H4, sources, sampling and word lists were committed to the public repository before any page was measured (docs/research/sgr-007-preregistration.md). Two clarifications made before measuring are listed there with reasons.
• Scientific abstracts: a seeded random sample of 10,000 from the 1,057,471 arXiv abstracts dated 2015 to November 2022 with no revision after ChatGPT's launch, from SGR-008's census of arXiv, cleaned of LaTeX and measured like every other text; the share of all eligible abstracts containing each word is reported alongside.
• Web pages: looked up in Common Crawl's own index files (read directly, not through its busy query server), fetched from its WARC archives and reduced to article text with the product's reader.
• Measurement: each page or passage is scored with the same deterministic code ScriptGrain uses, for the word lists, the sentence-shape scan, em dashes, sentence length and contractions. No language model reads or scores any text.
• Word lists: SGR-008's lists of 2024 tells and publicly mocked ChatGPT words (built from published studies of scientific writing) and SGR-004's AI-era marketing words, used unchanged.
• Tests: differences in mean rate per page with 95% bootstrap intervals (10,000 resamples, seed 20261004). Exploratory analyses are labelled as such.

What was measured

• AI-ism rate per 1,000 words
• Rate for each word list
• Delve rate
• AI sentence shapes per 1,000 words
• Em dashes per 1,000 words
• Average sentence length
• Contractions per 1,000 words
• Google Books frequency per million words, 1800 to 2019

Findings

• H1 (supported): press releases were the densest pre-AI home of AI-isms: 1.61 per 1,000 words, against 0.44 in news and 0.24 in Wikipedia. Press releases led every kind of writing measured, followed by company homepages (1.36). As pre-registered, press releases, marketing and SEO blogs (0.69) and essay-mill essays (0.53) each used more than news, all countries pooled, and more than Wikipedia. Press releases against news: difference 1.17 (95% interval 0.99 to 1.35). Essay mills only just cleared news: 0.09 (95% interval 0.02 to 0.17).
• Exploratory: the classic ChatGPT words were most at home in press releases: 14 of 44 words. For each word with at least 20 uses, we found the kind of writing where its rate was highest relative to everything measured. Of 44 words common enough to place, press releases were home to 14: "innovative", "advancements", "pivotal", "delve", "meticulous", "underscore", "comprehensive", "potential", "seamless", "empower", "streamline", "revolutionize", "reimagine", "redefine". Essay mills were home to "utilized", "firstly", the vocabulary of the student essay. The word lists were picked from what models over-use, so this is a description of where they lived, not a test.
• Exploratory: AI sentence shapes had a different, older home: 2.07 per 1,000 words in pre-1928 speeches, 0.45 in UK news. The shapes the AI cliché scan looks for ("not X but Y" balancing, reflexive lists of three, stock connectives) were densest not on the modern web but in pre-1928 speeches, sermons and religion, conduct of life and self-help, science. 7 of 8 book genres used them more often than UK news, US news and Wikipedia; the exception was fiction. Vocabulary and shape point to different ancestors: the announcement and the pulpit.
• Exploratory: scientific abstracts had the quieter words, not the famous ones: 0.36 AI-isms per 1,000 words (13 of 26); first of 26 for the quieter words. Pre-ChatGPT arXiv abstracts were mid-table on the main AI-ism rate but used the quieter LLM-favoured words more than any other writing measured (2.24 per 1,000 words of words such as "significant", "enhance", "additionally" and "potential"), and were home to "facilitates", "enhance", "additionally", "significant". Their AI sentence-shape rate (1.25) was the highest of any modern writing measured. The famous words were rare: across all 1,057,471 eligible abstracts, "delve" appeared in about 1 in 5,287. Studies of scientific writing since 2023 report large rises in exactly those words; see the citations.
• H2 (not supported): no clear 'delve' gap between Nigerian or Kenyan news and UK or US news: difference 0.005 per 1,000 words, 95% interval 0.000 to 0.013. "Delve" was used a little more in Nigerian and Kenyan news (pooled) than in UK and US news, in the predicted direction, but the interval reaches zero: it appeared on 3 of 600 Nigerian pages, 1 of 600 Kenyan pages and none of 1,000 UK and US pages. The wider 2024 tell list showed no difference either: 0.016 (95% interval -0.032 to 0.060). In public news writing from 2021, "delve" was too rare anywhere to carry the explanation on its own. This does not test the raters themselves.
• H4 (supported): the words were rising in books for decades before ChatGPT: 40 of 43 words more common in 2019 than in 1980. In Google Books, "innovative" took off around 1975 (0.14 uses per million words in 1950, 12.12 in 1980), "transformative" rose 31 times between 1980 and 2019, "reimagine" 106 times, and "delve" took off around 2007. The exceptions, "utilized", "commendable", "streamline", had peaked earlier.
• H3 (not supported): fiction was not the least AI-like writing; casual Q&A was: Q&A forum 0.04, pre-1928 fiction 0.19 per 1,000 words. We predicted that pre-1928 fiction would use AI-isms least. It was low, but several book genres were lower, and the lowest of all was English Stack Exchange, where people answer each other's questions in plain words. In this sample, writing that announces or persuades used the most AI-isms, and plain answers to questions used the fewest.

Limitations

• One snapshot of the web (October 2021) and a few sources per kind of writing. A different month or different outlets would move individual numbers; the large gaps, such as press releases against news, are the finding, not the decimals.
• Word lists were taken from studies of what models over-use, so finding them in press releases describes where those words lived; it does not show that press releases were over-represented in any model's training data.
• Pre-1928 books are a different era as well as a different genre, so their high rate of AI sentence shapes may reflect period style as much as genre.
• The sentence-shape scan is ScriptGrain's own instrument; it over-fires on some ordinary sentences, and another checker would count differently.
• Some 'homes' rest on a few dozen uses. The table shows the counts.

Read the full study

SGR-008: The AI tells moved: 2.1 million arXiv abstracts, 2015 to 2026

When words like "delve" became a joke, did language models stop shaping scientific writing, or did the tells change?

Every arXiv abstract from January 2015 to September 2026 (2,122,337 papers), measured against word lists fixed in advance. The 2024 AI tells peaked and fell below their pre-ChatGPT level; a different set of words rose in every field, so lists built on 2024 models now miss most of the signal.

Conducted 2026-10-04, published 2026-10-04, corrected 2026-10-12.

Sample

Every paper on arXiv, harvested on 4 October 2026 from arXiv's OAI-PMH service in the arXivRaw format, which carries the date of every version. 2,122,337 abstracts met the rules below. arXiv serves the latest version's abstract, which is why revision dates matter. (n = 2122337)

Method

• Pilot, 4 October 2026: Firecrawl's Research Index returned 41,657 arXiv abstracts for 38 categories x 3 fixed topics x 2015 to 2026. The pilot produced the 2026 word set and the idea; it is not part of any test.
• Pre-registration: hypotheses H1 to H5, the frozen word sets, the excess-word method, thresholds and exclusions were committed to the ScriptGrain repository before the census was analysed. Seven clarifications and one deviation followed, each before the census was measured and each listed with its reason.
• Census: every arXiv record from arXiv's OAI-PMH service (3,195,084 records), filtered by the rules above.
• Frozen-set tests: share of abstracts with at least one word of each set, 2020 to 2022 against 2024 and 2026 (January to September), with 95% confidence intervals from 2,000 bootstrap resamples of subject categories.
• Excess words: for 2023 to 2026, each word's share against a straight-line extension of its 2021 to 2022 trend. Words were chosen on papers submitted in odd months and scored on papers submitted in even months, so no word is scored on the papers that selected it. Words already in 5% or more of abstracts were left out of the word sets (the bound becomes unstable).
• Labelling: the 351 top candidates were labelled content or style by two independent model passes (agreement 343/351, Cohen's kappa 0.892) and a third pass for the 8 disagreements.
• Instruments: average sentence length, punctuation and the AI-tell scan were measured on every abstract with the code ScriptGrain runs (measureTextFeatures, scanCliches).
• Placebo (added after the confirmatory run): the lower-bound method was run five years earlier, 2016 to 2021, where no language models existed, to measure how much excess ordinary drift produces.

What was measured

• Share of abstracts using each word and word set
• Excess word share against trend
• Lower bound on model-processed abstracts
• Average sentence length
• AI-tell density per 1,000 words
• Em dashes per 1,000 words
• Words per abstract

Findings

• The 2024 tells rose, then fell below their pre-ChatGPT level: 4.5% to 11.9% to 3.0% of abstracts. The twelve best-known 2024 tells were in 4.54% of abstracts from 2020 to 2022, 11.85% in 2024 and 3.03% in 2026 (2024 minus baseline +5.4 to +8.8 points, 2026 minus 2024 -11.0 to -6.3 points, 95% intervals). 2026 is 1.5 points below the baseline (interval -2.1 to -0.9). By the second half of 2026 the set was at 2.0%, about its 2016 level. "Delve" went from 0.04% of abstracts in 2022 to 0.35% in 2024 and 0.02% in 2026.
• A new set of words rose in every field: 1.9% to 12.4% of abstracts. Twelve words found in the pilot (underexplored, grounded, aligns, synthesizes, replaces, leaving, recovers, expose, structurally, reproducible, treats, vanishes) were in 1.90% of abstracts from 2020 to 2022, 3.31% in 2024 and 12.39% in 2026 (2026 minus 2024 +6.8 to +10.9 points). Tested only on papers outside the pilot, the rise held in all eight subject groups: computer science 7.6 times, electrical engineering and systems 5.7 times, statistics 5.2 times, physics and astronomy 3.7 times, mathematics 3.5 times, quantitative biology 6.9 times, quantitative finance 6.8 times, economics 7.1 times. The two sets crossed in the second half of 2025 and the new set has led in every half-year since.
• The words that rose most in 2024 and in 2026 are different words: 6 of the top 50 style words in common (Jaccard 0.06). Ranking every style word by how far it rose above its own trend, the top 50 for 2024 and the top 50 for 2026 share six words: critical, distinct, enabling, findings, practical, remains. The 2024 list leads with challenges, additionally, findings, enhance, insights, crucial, comprehensive, enhancing; the 2026 list with remains, yet, rather, establish, explicit, enabling, every, remain. Both lists are below in full.
• Lists built on 2024 vocabulary now undercount: at least 68% (2026 words) against 19% (2024 words) of 2026 abstracts. The same lower-bound method, on the same 2026 abstracts, gives at least 68% (interval 61 to 73%) with the 50 style words that rose most in 2026 and 19% (interval 9 to 26%) with the 50 that rose most in 2024. With 2024 words the bound peaks in 2025 (39%) and falls; with 2026 words it rises every year: 2023 12%, 2024 22%, 2025 47%, 2026 68%. A placebo run on 2016 to 2021, before any language model, gave 0% for the 2026 words and 10% for the 2024 words, which were already drifting upward (Study SGR-007 traces where those habits came from); re-running the whole selection on pre-LLM years produced up to 10% (interval up to 19%). Read the bounds with that margin in mind.
• Sentence length did not move: 24.5 to 24.5 words per sentence. Average sentence length was 24.46 words in 2020 to 2022 and 24.49 in 2026 (difference +0.03; 90% interval -0.11 to +0.17, inside the pre-registered one-word margin). The change is in vocabulary, not rhythm.
• No sign that old abstracts were re-polished: 4.4% against 4.5% (2024 tells, revised against unrevised). 11.2% of 2020 to 2022 papers have a version dated after November 2022, so their current abstract may have been edited since. They carry the 2024 tells at 4.39% against 4.54% for papers never revised, and the 2026 words at 2.27% against 1.90%. Leaving them in or out changes no result (sensitivity table below).

Limitations

• arXiv only. It is mostly physics, mathematics and computer science, and biomedicine is thin; the published PubMed studies cover that ground.
• Word shares describe a body of writing. They cannot say which papers used a model, and people also pick up model vocabulary by reading it.
• The topic mix changed: far more 2026 papers are about language models and agents. Style words were separated from topic words by labelling, and the new set rose in mathematics and physics too, but some topic effect may remain in the bounds.
• The lower bounds assume each word's 2021 to 2022 trend would have continued in a straight line. The placebo shows ordinary drift can add about 10 points over four years, more for the 2024 words, which were already rising.
• arXiv serves each paper's latest abstract. Baseline papers revised after November 2022 were removed; for 2024 and 2026 the abstract may postdate first submission (31% of 2024 papers have a version more than 90 days after the first).
• The 2026 word set came from a pilot sample chosen by semantic search, not at random. That is why the census test excludes the pilot papers.
• Content or style labels were assigned by Claude models (agreement 343 of 351 words). A different labeller would move a few words between the lists.
• 2026 covers January to September only.

Read the full study

SGR-009: Do brands in the same sector sound the same? 89 UK brands' websites, measured

Measured with a pairwise voice score, are two pages from the same brand's website more alike than a page from that brand and a page from a rival in the same sector?

A pre-registered study of 299 pages from 89 UK brands' websites in 10 sectors, from banks to charities: homepage, about, product, pricing and careers pages, every pair scored with ScriptGrain's pairwise voice score and judged by two AI models in five runs. A brand's own pages were barely more alike than a rival's. Free dataset under CC BY 4.0.

Conducted 2026-10-04, published 2026-10-04.

Sample

A panel of 96 UK consumer brands in 10 sectors (10 per sector, 6 in energy), frozen before fetching, none from the pilot and no two in a sector sharing a parent company. For each brand: the homepage plus the shortest URL matching fixed rules for an about, product, pricing and careers page, fetched on 4 October 2026 from a UK location and reduced to text with the product's own page reader, cut to the first 700 words. 89 brands had at least two readable pages. (n = 299)

Method

• Pre-registration: hypotheses H1 to H3, the brand panel, the page-type rules and both judges were committed to the public repository before any page was fetched (docs/research/sgr-009-preregistration.md and scripts/experiments/sgr-009/panel.json).
• Pages: Firecrawl's site map of each brand, the page-type rules applied to its links (shortest URL first, up to three tries per type), Firecrawl's raw HTML from a UK location, and ScriptGrain's own page reader, exactly what the product reads when given a URL.
• Judges: every page judged with the production judge prompt. Primary: Claude Haiku 4.5 at temperature 0 with the production limits, run twice. Robustness: Claude Sonnet 5.5, run three times (its temperature cannot be set). No judge call failed.
• Scoring: every ordered pair of included pages through the production computation, once per judge run and once with counted features only.
• Tests: AUC and difference in medians with 95% intervals from 2,000 bootstrap resamples of brands within sectors (fixed seed). Analyses beyond H1 to H3 are labelled exploratory.

What was measured

• Pairwise voice score (raw and band)
• AUC, same brand against rival brands
• Counted features per page
• Judged attributes per page, five runs
• Judge agreement between runs

Findings

• H1 (supported): a brand's own pages were barely more alike than a rival's: AUC 0.551 (95% interval 0.521 to 0.582). We predicted the score would tell a brand's own page pairs from rival pairs in the same sector less than 60% of the time. It did so 55% of the time, in line with the pilot (0.54). The interval stays above 0.50, so there is a little brand signal, but it is small: the middle of the range for a brand's own pairs (median raw 0.654) sits close to rivals' (0.629). Against brands in other sectors the figure was 0.610.
• H2 (not supported): page type did not beat brand: median 0.657 (rivals, same page type) against 0.654 (same brand, different page type); difference 0.003 (95% interval -0.012 to 0.028). We predicted, from the pilot, that two rivals' pages of the same type would score closer than two different pages from one brand. They scored the same. The pilot's version of the test also reversed: rival homepages scored a median 0.649 against each other and a homepage scored 0.657 against its own inner pages (pilot: 0.673 and 0.653). Page type does carry some signal, about as much as brand: among pairs from different brands, same-type pairs scored higher than different-type pairs 58% of the time.
• H3 (supported): sector carried a little signal: AUC 0.563 (95% interval 0.542 to 0.586). A rival's page in the same sector scored higher than a page from another sector 56% of the time. 33% of other-sector pairs read off voice, against 26% of same-sector pairs and 24% of a brand's own.
• Robustness: every judge gave the same answer: H1 AUC 0.558, 0.564, 0.558 on three Sonnet 5.5 runs; 0.553 with no judge. As pre-registered, the result counts as robust to the judge only if H1 held on all three Sonnet runs and H2 and H3 matched on at least two. All three runs matched on all three hypotheses. The two models read the pages differently (they agreed exactly on 55% of attribute readings), yet the conclusions did not move. Haiku at temperature 0 was close to repeatable (98.4% of readings identical across two runs; median change in a pair's raw score 0, largest 0.114); Sonnet's three runs agreed on 89%.
• Exploratory: sectors differ, but no sector stands out clearly: AUC from 0.66 (energy) to 0.49 (banks). Energy and travel brands were the easiest to tell from their rivals; banks, software and insurance the hardest. Each sector has six to ten brands, so these are point estimates without intervals and small differences between sectors should not be read as rankings. By page type the AUC ranged from 0.53 to 0.57; using only pages of 600 words or more gave 0.555.

Limitations

• 299 pages from 89 brands rather than the 480 the design allowed: careers and pricing pages were often missing, and some homepages build their text in the browser. The rules were applied as frozen and every gap is shown above.
• The page reader keeps navigation, footer and cookie text, which every page of one site shares. That raises same-brand scores, so the true separation of brand voice by this measure is, if anything, lower.
• Website pages are written by many people and agencies over years. 'Same brand' is a label, not a single writer.
• UK consumer brands in one week of October 2026, in English. Business-to-business, public-sector and personal sites are not covered.
• Both judges are Claude models; another vendor's model would read the pages differently. The counted-features-only result does not depend on any model.
• Per-sector figures rest on six to ten brands each and have no intervals.

Read the full study

SGR-010: How do UK brands write? 16,070 blog posts and press releases, measured

What is normal in the long-form writing of UK brands, by sector and by channel, and does a brand write the same way in its press releases as on its blog?

A baseline for brand writing: 16,070 blog posts, press releases, announcements and reports from 117 UK brands in 10 sectors, 2015 to 2026, measured with ScriptGrain's instruments. Typical ranges for sentence length, "you", contractions, dashes and AI-associated words by sector and channel, two pre-registered comparisons, and a free dataset under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-05.

Sample

The brand deep corpus (October 2026): 117 UK consumer brands in 10 sectors (banks, energy, telecoms, software, supermarkets, travel, insurance, fashion, charities, food and drink). For each, blog posts, press releases, product announcements and information packs found on its own sites by fixed URL rules and fetched on 4 and 5 October 2026, plus archived pages from 2015 to November 2022. 11,742 live and 4,328 archived pages were measured. (n = 16070)

Method

• Corpus: the brand deep corpus (October 2026): blog posts, press releases, product announcements and information packs from 122 UK consumer brands in 10 sectors, sampled by fixed rules from each brand's own sites, plus 4,845 archived pages from before ChatGPT's launch read from the Wayback Machine as they were published. Scripts and the full method: scripts/experiments/brand-corpus/ in the ScriptGrain repository.
• Text measured: each page's main content (Firecrawl markdown) reduced to plain text, so menus and footers do not count. Pages under 150 words, exact duplicates within a brand, and pages where more than 40% of the words sit in lines of six words or fewer (link lists, menus) were excluded.
• Measures: computed in code by ScriptGrain's own instruments, the same counts the free tools run. No language model read or scored any page.
• Pre-registration: the hypotheses, eras, exclusions and tests for this series were committed to the public repository before any page was measured (docs/research/brand-series-preregistration.md). One rule was changed before any result was looked at, and the file records why.

What was measured

• Average sentence length
• Sentence-length variation
• "You", "we" and "I" shares
• Contractions per 1,000 words
• Commas per sentence
• Questions and exclamation marks per 1,000 words
• Long-word share
• Em dashes per 1,000 words
• AI-associated words per 1,000 words
• AI sentence shapes per 1,000 words

Findings

• O1 (supported): press releases use longer sentences than the same brand's blog: +2.9 (95% interval 2.2 to 3.5) words per sentence. In 53 brands with at least 10 press releases and 10 blog posts each, the press release sentence was longer in the brand average. Press releases also used 12% more long words and 39% more commas per sentence.
• O2 (supported): blogs speak to "you" more than the same brand's press releases: +30 points of the I, we and you mix (95% interval 25 to 35). Press releases are written about the company ("we", the brand name) and blogs to the reader. Questions follow the same split: blog posts asked 3.8 times as many questions as press releases.
• Exploratory: AI-associated words are more common in press releases: 1.52 per 1,000 words in press releases, 1.02 on blogs. Within the same brands, press releases used 37% more of the listed words (95% interval 13% to 61%). That matches SGR-007, which found press releases were the densest home of these words before ChatGPT existed.
• Exploratory: sectors write differently: "You" from 20% (supermarkets) to 66% (insurance); sentences from 10.0 to 14.8 words. Sectors that sell to individuals over the phone and online (insurance, software, telecoms) wrote most to "you". Supermarkets and charities wrote most about themselves and their work. Food and drink brands used the most contractions (22.2 per 1,000 words). Sector figures rest on 8 to 18 brands each and carry no intervals.

Limitations

• Pages were chosen by fixed URL rules from what each brand publishes, not at random, and brands publish very different amounts.
• Ten sectors of UK consumer brands; business-to-business brands, public bodies and other countries are not covered.
• Every measure is a count of surface features. None says whether writing is good, effective or on-brand.
• Sector figures rest on 8 to 18 brands each.

Read the full study

SGR-011: Did brand writing change after ChatGPT? 37 UK brands, before and after

Within the same brands, did blog posts and press releases written after ChatGPT use more of the words and habits associated with AI writing than those written before it?

A pre-registered before-and-after study of UK brands' own writing: each brand's blog posts and press releases from 2019 to 2022, read as published from the Wayback Machine, against the same brand's writing from 2024 to 2026. AI-associated words rose 40%; em dashes, AI sentence shapes and sentence length did not change. Free dataset under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-05.

Sample

UK consumer brands from the brand deep corpus (October 2026) with at least 10 dated pages in the same channel both before (2019 to November 2022) and after (2024 to 2026): 37 brands and 49 brand-channel pairs. Before pages are mostly the archived originals from the Wayback Machine; after pages come from the brands' live sites, fetched on 4 and 5 October 2026. (n = 49)

Method

• Corpus: the brand deep corpus (October 2026): blog posts, press releases, product announcements and information packs from 122 UK consumer brands in 10 sectors, sampled by fixed rules from each brand's own sites, plus 4,845 archived pages from before ChatGPT's launch read from the Wayback Machine as they were published. Scripts and the full method: scripts/experiments/brand-corpus/ in the ScriptGrain repository.
• Text measured: each page's main content (Firecrawl markdown) reduced to plain text, so menus and footers do not count. Pages under 150 words, exact duplicates within a brand, and pages where more than 40% of the words sit in lines of six words or fewer (link lists, menus) were excluded.
• Measures: computed in code by ScriptGrain's own instruments, the same counts the free tools run. No language model read or scored any page.
• Pre-registration: the hypotheses, eras, exclusions and tests for this series were committed to the public repository before any page was measured (docs/research/brand-series-preregistration.md). One rule was changed before any result was looked at, and the file records why.
• Tests: within-brand change in brand-weighted means, with 95% intervals from 2,000 bootstrap resamples of brands within sectors (fixed seed). H2 compares the per-year rate of change from 2015-18 to 2019-22 with that from 2019-22 to 2024-26, for brands with all three eras. Analyses beyond H1 to H5 are labelled exploratory.

What was measured

• AI-associated words per 1,000 words (and each word list)
• AI sentence shapes per 1,000 words
• Em dashes per 1,000 words
• Average sentence length
• Contractions per 1,000 words
• "You" share

Findings

• H1 (supported): AI-associated words rose within the same brands: +0.41 (95% interval 0.16 to 0.65) per 1,000 words, +40%. From 1.03 to 1.44 per 1,000 words across 37 brands. By word list: AI-era marketing words +45% (interval clears zero), the quieter LLM-favoured words +74% (clears zero), the 2024 tells such as "delve" +20% and the publicly mocked words +22% (neither clears zero).
• H2 (not supported): the rise was not clearly a break from the trend before: 0.086 per 1,000 words a year after, against 0.060 a year from 2015-18 to 2019-22. For the 22 brands with pages in all three eras, the yearly rise after ChatGPT was larger, but the difference (+0.026, 95% interval −0.049 to 0.101) includes zero. These words were already climbing in brand writing before ChatGPT, as SGR-007 found in books and the web. The calendar-year series below shows a sharper jump in 2023 and 2024 than this per-year test can resolve.
• H3 (not supported): em dashes did not rise: 2.43 to 2.50 per 1,000 words; change +0.07 (95% interval −0.36 to 0.49). Despite the 'ChatGPT dash' reputation, brand writing used em dashes at the same rate after ChatGPT as before. Exploratory: press releases did gain dashes (+0.52, 95% interval 0.11 to 0.91), blog posts did not (−0.15, 95% interval −0.67 to 0.37).
• H4 (not supported): AI sentence shapes did not rise: 0.24 to 0.26 per 1,000 words; change +0.02 (95% interval −0.03 to 0.07). 'Not X but Y' constructions, reflexive lists of three and the cliché checker's stock phrases were as rare after ChatGPT as before.
• H5 (supported): sentence length stayed within 1.5 words: 14.7 to 14.3 words; change −0.43 (95% interval −0.95 to 0.12). The 90% interval (−0.88 to 0.02) sits inside ±1.5 words. Sentences got slightly shorter, if anything.
• Robustness: the result held under every pre-registered check: H1 with text dates added +0.53 (95% interval 0.32 to 0.73); with 2023 counted as after +0.37 (95% interval 0.13 to 0.60). Archived and live copies of 2019-22 pages (39 brands with both) used AI-associated words at the same rate (difference +0.10, 95% interval −0.15 to 0.32), so mixing archived before-pages with live after-pages does not create the rise. Live copies did carry fewer em dashes (−0.39, 95% interval −0.70 to −0.08) and slightly longer sentences than their archived versions, a sign some old pages were edited since.

Limitations

• Before pages are mostly archived copies and after pages are live. The source check above found no difference in AI-associated words between archived and live copies of the same era, but live copies had fewer em dashes, so the em dash comparison may understate any rise.
• A page's date is its publication date where the page says so, and otherwise the month it was first archived (a latest possible date).
• Word lists count spellings. These words rose in human writing for decades before ChatGPT, and topics shifted too: brands now write more about AI itself.
• Brands publish different amounts, and only brands with enough pages in both eras are in the test, which favours brands with active blogs and newsrooms.
• UK consumer brands only.

Read the full study

SGR-012: Does a brand keep one voice across its press releases and blog? 14 UK brands, measured

Measured with a writing-voice score, are two pages from the same brand and channel more alike than the same brand's blog and press releases, and is a rival's page in the same channel as close as the brand's own other channel?

A pre-registered test on 1,074 blog posts and press releases from 14 UK brands, 2024 to 2026: 211,262 page pairs scored with ScriptGrain's counted-features voice score. A brand's pages in one channel agreed far more than its pages across channels, and a rival's page in the same channel came about as close as the brand's own other channel. Free data under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-05.

Sample

UK consumer brands from ScriptGrain's brand deep corpus (October 2026) with at least 8 blog posts and 8 press releases or announcements of 300 words or more dated January 2024 or later: 14 brands, 1,074 pages, 211,262 scored pairs. (n = 1074)

Method

• Pre-registration: hypotheses C1 and C2 were committed with the series plan before any page was measured (docs/research/brand-series-preregistration.md section 5), and four clarifications (how channels are pooled, which dates count, how medians are taken) before any pair was scored.
• Text: each page's main content reduced to plain text, as in SGR-010 and SGR-011.
• Scoring: the production computeVoiceMatch with pseudoProfileFromText and no AI judge (counted features only), the computation behind the free checker and the compare-voice API. SGR-009 found counted features alone carry as much brand signal as the full score on web copy.
• Tests: C1 and C2 are the within-brand difference in median scores, averaged over brands, with a 95% brand bootstrap within sectors.

What was measured

• Pairwise voice score (raw and display)
• Pair kind
• Brand, sector, channel

Findings

• C1 (supported): a brand's pages agree more within a channel than across channels: +0.080 (95% interval 0.055 to 0.105) on the raw scale; 81 against 68 out of 100. The gap held in 13 of 14 brands. On the checker's bands, same-channel pairs sit at the "on voice" line and cross-channel pairs in "drifting".
• C2 (not supported): a rival's same-channel page came about as close as the brand's own other channel: +0.017 (95% interval −0.010 to 0.041) on the raw scale; 71 against 68 out of 100. We predicted, as SGR-009 had for landing pages, that channel would beat brand: that a rival's blog would score closer to a brand's blog than the brand's own press releases do. On long-form writing the two were level overall (7 of 13 brands leaned that way), so the hypothesis is not supported. Channel and brand pulled about equally.
• Exploratory: sectors split: Channel ahead in software, insurance; brand ahead in energy, charities. Each sector here has two to five brands, so these are descriptions, not findings.

Limitations

• 14 brands qualified: many brands publish too few press releases or too few long blog posts. Sector rows rest on one to five brands.
• The score counts surface habits only. Two pages can agree on sentence length and punctuation and still read differently to a person.
• Recent writing only (2024 to 2026).
• UK consumer brands only.

Read the full study

SGR-013: How do UK banks and building societies write? 18 brands measured against nine other sectors

How does the writing of UK banks and building societies differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?

2,669 blog posts and press releases from 18 UK banks and building societies, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-06.

Sample

18 UK banks and building societies from ScriptGrain's brand deep corpus (October 2026): blog posts, press releases, product announcements and information packs from their own sites, fetched on 4 and 5 October 2026, plus archived pages from 2015 to November 2022. 2,669 pages measured. (n = 2669)

Method

• Plan: the sector reports were set out in the series pre-registration (docs/research/brand-series-preregistration.md section 6) before any page was measured: the SGR-010 tables for the sector against all other sectors, and the SGR-011 measures within the sector, exploratory, with intervals.
• Text and measures: as in SGR-010. Each page's main content reduced to plain text and counted in code.
• Weighting: each brand's pages are averaged first, then brands, here and for the other nine sectors.
• Years: pages with a date from their metadata, structured data or URL, or an archive capture date. A year is shown for the sector when it has 10 or more pages.
• Before and after: SGR-011's rule, inside the sector: a brand channel with at least 10 dated pages from 2019 to November 2022 and 10 from 2024 on.

What was measured

• Words per sentence
• Sentence length variation
• Long words (7+ letters)
• "You" share of I, we, you
• "We" share of I, we, you
• Contractions per 1,000 words
• Questions per 1,000 words
• Exclamation marks per 1,000 words
• Em dashes per 1,000 words
• AI-associated words per 1,000
• AI-era marketing words per 1,000
• Words per page

Findings

• Fewer AI-era marketing words than other sectors: 0.71 against 0.94 (−25%). Banks ranked 4th lowest of 10 sectors. The middle half of the sector's brands ran from 0.30 to 0.94.
• Fewer AI-associated words than other sectors: 0.96 against 1.19 (−19%). Banks ranked 4th lowest of 10 sectors. The middle half of the sector's brands ran from 0.53 to 1.33.
• AI-associated words by year: Peak in 2024: 2.20 per 1,000 words. Across all 10 sectors the peak was 2024 (2.16). Banks ranked 4th lowest of 10 sectors on these words over the whole period.
• Press releases against the same brands' blogs: Longer sentences (+2.2 words, 95% interval 1.3 to 3.2); less "you" (−32.4 points, 95% interval −44.9 to −20.2); fewer contractions (−4.3, 95% interval −6.9 to −1.6); more AI-associated words (+0.72, 95% interval 0.11 to 1.42). 13 banks had at least five of each. Positive numbers mean the press releases had more.
• Before and after ChatGPT (exploratory): 0.94 to 1.54 AI-associated words per 1,000; three brands. Three of the 18 banks had at least 10 dated pages in one channel both before ChatGPT (2019 to November 2022) and from 2024. In that writing, AI-associated words went from 0.94 to 1.54 per 1,000 words (+0.61, 95% interval 0.49 to 0.80). With three brands, treat that as a description of those brands, not of the sector.

Limitations

• 18 brands. A sector average can move with one or two brands; the brand range column shows the spread.
• Pages are what each brand publishes on its own sites. Brands that publish little have few pages.
• Counts of surface habits only. They describe style, not quality.
• Descriptive. Differences from other sectors are reported without significance tests.

Read the full study

SGR-014: How do UK energy suppliers and energy brands write? 11 brands measured against nine other sectors

How does the writing of UK energy suppliers and energy brands differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?

1,949 blog posts and press releases from 11 UK energy suppliers and energy brands, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-07.

Sample

11 UK energy suppliers and energy brands from ScriptGrain's brand deep corpus (October 2026): blog posts, press releases, product announcements and information packs from their own sites, fetched on 4 and 5 October 2026, plus archived pages from 2015 to November 2022. 1,949 pages measured. (n = 1949)

Method

• Plan: the sector reports were set out in the series pre-registration (docs/research/brand-series-preregistration.md section 6) before any page was measured: the SGR-010 tables for the sector against all other sectors, and the SGR-011 measures within the sector, exploratory, with intervals.
• Text and measures: as in SGR-010. Each page's main content reduced to plain text and counted in code.
• Weighting: each brand's pages are averaged first, then brands, here and for the other nine sectors.
• Years: pages with a date from their metadata, structured data or URL, or an archive capture date. A year is shown for the sector when it has 10 or more pages.
• Before and after: SGR-011's rule, inside the sector: a brand channel with at least 10 dated pages from 2019 to November 2022 and 10 from 2024 on.

What was measured

• Words per sentence
• Sentence length variation
• Long words (7+ letters)
• "You" share of I, we, you
• "We" share of I, we, you
• Contractions per 1,000 words
• Questions per 1,000 words
• Exclamation marks per 1,000 words
• Em dashes per 1,000 words
• AI-associated words per 1,000
• AI-era marketing words per 1,000
• Words per page

Findings

• More em dashes than other sectors: 2.96 against 2.21 (+34%). Energy companies ranked 2nd highest of 10 sectors. The middle half of the sector's brands ran from 1.61 to 3.77.
• More questions than other sectors: 3.9 against 3.0 (+30%). Energy companies ranked 3rd highest of 10 sectors. The middle half of the sector's brands ran from 2.1 to 5.0.
• Longer sentences than other sectors: 14.8 against 13.1 (+13%). Energy companies ranked highest of 10 sectors. The middle half of the sector's brands ran from 14.1 to 15.7.
• More even sentence lengths than other sectors: 0.70 against 0.79 (−10%). Energy companies ranked lowest of 10 sectors. The middle half of the sector's brands ran from 0.61 to 0.75.
• AI-associated words by year: Peak in 2024: 2.07 per 1,000 words. Across all 10 sectors the peak was 2024 (2.16). Energy companies ranked 3rd highest of 10 sectors on these words over the whole period.
• Press releases against the same brands' blogs: Less "you" (−19.6 points, 95% interval −28.1 to −11.0). Nine energy companies had at least five of each. Positive numbers mean the press releases had more.
• Before and after ChatGPT (exploratory): 1.09 to 1.49 AI-associated words per 1,000; four brands. Four of the 11 energy companies had at least 10 dated pages in one channel both before ChatGPT (2019 to November 2022) and from 2024. In that writing, AI-associated words went from 1.09 to 1.49 per 1,000 words (+0.40, 95% interval −0.15 to 1.15). With four brands, treat that as a description of those brands, not of the sector.
• One voice across channels (from SGR-012): Same brand, same channel 84; same brand, other channel 79; rival, same channel 70 (out of 100). Three energy companies qualified for SGR-012. Here the brand's own other channel scored closer than a rival's page in the same channel.

Limitations

• 11 brands. A sector average can move with one or two brands; the brand range column shows the spread.
• Pages are what each brand publishes on its own sites. Brands that publish little have few pages.
• Counts of surface habits only. They describe style, not quality.
• Descriptive. Differences from other sectors are reported without significance tests.

Read the full study

SGR-015: How do UK mobile, broadband and TV providers write? 15 brands measured against nine other sectors

How does the writing of UK mobile, broadband and TV providers differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?

1,816 blog posts and press releases from 15 UK mobile, broadband and TV providers, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-08.

Sample

15 UK mobile, broadband and TV providers from ScriptGrain's brand deep corpus (October 2026): blog posts, press releases, product announcements and information packs from their own sites, fetched on 4 and 5 October 2026, plus archived pages from 2015 to November 2022. 1,816 pages measured. (n = 1816)

Method

• Plan: the sector reports were set out in the series pre-registration (docs/research/brand-series-preregistration.md section 6) before any page was measured: the SGR-010 tables for the sector against all other sectors, and the SGR-011 measures within the sector, exploratory, with intervals.
• Text and measures: as in SGR-010. Each page's main content reduced to plain text and counted in code.
• Weighting: each brand's pages are averaged first, then brands, here and for the other nine sectors.
• Years: pages with a date from their metadata, structured data or URL, or an archive capture date. A year is shown for the sector when it has 10 or more pages.
• Before and after: SGR-011's rule, inside the sector: a brand channel with at least 10 dated pages from 2019 to November 2022 and 10 from 2024 on.

What was measured

• Words per sentence
• Sentence length variation
• Long words (7+ letters)
• "You" share of I, we, you
• "We" share of I, we, you
• Contractions per 1,000 words
• Questions per 1,000 words
• Exclamation marks per 1,000 words
• Em dashes per 1,000 words
• AI-associated words per 1,000
• AI-era marketing words per 1,000
• Words per page

Findings

• More questions than other sectors: 4.2 against 2.9 (+44%). Telecoms brands ranked 2nd highest of 10 sectors. The middle half of the sector's brands ran from 2.9 to 5.3.
• More "you" than other sectors: 62% against 48% (+28%). Telecoms brands ranked 3rd highest of 10 sectors. The middle half of the sector's brands ran from 52% to 78%.
• More AI-era marketing words than other sectors: 1.06 against 0.88 (+20%). Telecoms brands ranked 2nd highest of 10 sectors. The middle half of the sector's brands ran from 0.63 to 1.35.
• More contractions than other sectors: 17.5 against 15.5 (+13%). Telecoms brands ranked 3rd highest of 10 sectors. The middle half of the sector's brands ran from 13.1 to 21.6.
• AI-associated words by year: Peak in 2024: 1.97 per 1,000 words. Across all 10 sectors the peak was 2024 (2.16). Telecoms brands ranked 2nd highest of 10 sectors on these words over the whole period.
• Press releases against the same brands' blogs: Less "you" (−16.9 points, 95% interval −34.8 to −1.0); fewer contractions (−4.0, 95% interval −7.4 to −0.7). Nine telecoms brands had at least five of each. Positive numbers mean the press releases had more.
• Before and after ChatGPT (exploratory): 0.88 to 1.32 AI-associated words per 1,000; four brands. Four of the 15 telecoms brands had at least 10 dated pages in one channel both before ChatGPT (2019 to November 2022) and from 2024. In that writing, AI-associated words went from 0.88 to 1.32 per 1,000 words (+0.44, 95% interval −0.53 to 1.33). With four brands, treat that as a description of those brands, not of the sector.

Limitations

• 15 brands. A sector average can move with one or two brands; the brand range column shows the spread.
• Pages are what each brand publishes on its own sites. Brands that publish little have few pages.
• Counts of surface habits only. They describe style, not quality.
• Descriptive. Differences from other sectors are reported without significance tests.

Read the full study

SGR-016: How do software companies with UK sites write? 16 brands measured against nine other sectors

How does the writing of software companies with UK sites differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?

3,422 blog posts and press releases from 16 software companies with UK sites, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-09.

Sample

16 software companies with UK sites from ScriptGrain's brand deep corpus (October 2026): blog posts, press releases, product announcements and information packs from their own sites, fetched on 4 and 5 October 2026, plus archived pages from 2015 to November 2022. 3,422 pages measured. (n = 3422)

Method

• Plan: the sector reports were set out in the series pre-registration (docs/research/brand-series-preregistration.md section 6) before any page was measured: the SGR-010 tables for the sector against all other sectors, and the SGR-011 measures within the sector, exploratory, with intervals.
• Text and measures: as in SGR-010. Each page's main content reduced to plain text and counted in code.
• Weighting: each brand's pages are averaged first, then brands, here and for the other nine sectors.
• Years: pages with a date from their metadata, structured data or URL, or an archive capture date. A year is shown for the sector when it has 10 or more pages.
• Before and after: SGR-011's rule, inside the sector: a brand channel with at least 10 dated pages from 2019 to November 2022 and 10 from 2024 on.

What was measured

• Words per sentence
• Sentence length variation
• Long words (7+ letters)
• "You" share of I, we, you
• "We" share of I, we, you
• Contractions per 1,000 words
• Questions per 1,000 words
• Exclamation marks per 1,000 words
• Em dashes per 1,000 words
• AI-associated words per 1,000
• AI-era marketing words per 1,000
• Words per page

Findings

• More AI-era marketing words than other sectors: 1.80 against 0.76 (+137%). Software companies ranked highest of 10 sectors. The middle half of the sector's brands ran from 1.21 to 2.39.
• More AI-associated words than other sectors: 2.12 against 1.01 (+111%). Software companies ranked highest of 10 sectors. The middle half of the sector's brands ran from 1.50 to 2.92.
• Longer pages than other sectors: 1,512 against 1,033 (+46%). Software companies ranked 2nd highest of 10 sectors. The middle half of the sector's brands ran from 1,057 to 1,654.
• More em dashes than other sectors: 3.04 against 2.16 (+41%). Software companies ranked highest of 10 sectors. The middle half of the sector's brands ran from 2.27 to 3.79.
• AI-associated words by year: Peak in 2024: 4.07 per 1,000 words. Across all 10 sectors the peak was 2024 (2.16). Software companies ranked highest of 10 sectors on these words over the whole period.
• Press releases against the same brands' blogs: Longer sentences (+1.4 words, 95% interval 0.3 to 2.4); less "you" (−25.6 points, 95% interval −35.5 to −15.1); more AI-associated words (+0.98, 95% interval 0.60 to 1.33). 15 software companies had at least five of each. Positive numbers mean the press releases had more.
• Before and after ChatGPT (exploratory): 1.84 to 2.41 AI-associated words per 1,000; ten brands. Ten of the 16 software companies had at least 10 dated pages in one channel both before ChatGPT (2019 to November 2022) and from 2024. In that writing, AI-associated words went from 1.84 to 2.41 per 1,000 words (+0.57, 95% interval −0.14 to 1.14).
• One voice across channels (from SGR-012): Same brand, same channel 79; same brand, other channel 66; rival, same channel 75 (out of 100). Five software companies qualified for SGR-012. Here a rival's page in the same channel scored closer than the brand's own other channel.

Limitations

• 16 brands. A sector average can move with one or two brands; the brand range column shows the spread.
• Pages are what each brand publishes on its own sites. Brands that publish little have few pages.
• Counts of surface habits only. They describe style, not quality.
• Descriptive. Differences from other sectors are reported without significance tests.

Read the full study

SGR-017: How do UK insurers and insurance brands write? 10 brands measured against nine other sectors

How does the writing of UK insurers and insurance brands differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?

2,049 blog posts and press releases from 10 UK insurers and insurance brands, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-10.

Sample

10 UK insurers and insurance brands from ScriptGrain's brand deep corpus (October 2026): blog posts, press releases, product announcements and information packs from their own sites, fetched on 4 and 5 October 2026, plus archived pages from 2015 to November 2022. 2,049 pages measured. (n = 2049)

Method

• Plan: the sector reports were set out in the series pre-registration (docs/research/brand-series-preregistration.md section 6) before any page was measured: the SGR-010 tables for the sector against all other sectors, and the SGR-011 measures within the sector, exploratory, with intervals.
• Text and measures: as in SGR-010. Each page's main content reduced to plain text and counted in code.
• Weighting: each brand's pages are averaged first, then brands, here and for the other nine sectors.
• Years: pages with a date from their metadata, structured data or URL, or an archive capture date. A year is shown for the sector when it has 10 or more pages.
• Before and after: SGR-011's rule, inside the sector: a brand channel with at least 10 dated pages from 2019 to November 2022 and 10 from 2024 on.

What was measured

• Words per sentence
• Sentence length variation
• Long words (7+ letters)
• "You" share of I, we, you
• "We" share of I, we, you
• Contractions per 1,000 words
• Questions per 1,000 words
• Exclamation marks per 1,000 words
• Em dashes per 1,000 words
• AI-associated words per 1,000
• AI-era marketing words per 1,000
• Words per page

Findings

• Fewer AI-era marketing words than other sectors: 0.47 against 0.95 (−50%). Insurers ranked 2nd lowest of 10 sectors. The middle half of the sector's brands ran from 0.31 to 0.53.
• Fewer AI-associated words than other sectors: 0.62 against 1.21 (−49%). Insurers ranked 2nd lowest of 10 sectors. The middle half of the sector's brands ran from 0.40 to 0.74.
• More questions than other sectors: 4.3 against 2.9 (+47%). Insurers ranked highest of 10 sectors. The middle half of the sector's brands ran from 3.5 to 4.9.
• More "you" than other sectors: 66% against 48% (+38%). Insurers ranked highest of 10 sectors. The middle half of the sector's brands ran from 55% to 82%.
• AI-associated words by year: Peak in 2024: 1.04 per 1,000 words. Across all 10 sectors the peak was 2024 (2.16). Insurers ranked 2nd lowest of 10 sectors on these words over the whole period.
• Press releases against the same brands' blogs: Less "you" (−23.3 points, 95% interval −37.0 to −7.6); fewer contractions (−7.0, 95% interval −9.5 to −4.6); more AI-associated words (+0.66, 95% interval 0.18 to 1.18). Seven insurers had at least five of each. Positive numbers mean the press releases had more.
• Before and after ChatGPT (exploratory): 0.60 to 0.81 AI-associated words per 1,000; seven brands. Seven of the 10 insurers had at least 10 dated pages in one channel both before ChatGPT (2019 to November 2022) and from 2024. In that writing, AI-associated words went from 0.60 to 0.81 per 1,000 words (+0.21, 95% interval −0.22 to 0.66).
• One voice across channels (from SGR-012): Same brand, same channel 81; same brand, other channel 63; rival, same channel 74 (out of 100). Three insurers qualified for SGR-012. Here a rival's page in the same channel scored closer than the brand's own other channel.

Limitations

• 10 brands. A sector average can move with one or two brands; the brand range column shows the spread.
• Pages are what each brand publishes on its own sites. Brands that publish little have few pages.
• Counts of surface habits only. They describe style, not quality.
• Descriptive. Differences from other sectors are reported without significance tests.

Read the full study

SGR-018: How do UK charities write? 10 brands measured against nine other sectors

How does the writing of UK charities differ from other sectors' on counted habits such as sentence length, "you", contractions and AI-associated words, and how has it changed since 2015?

1,994 blog posts and press releases from 10 UK charities, measured on the same counted habits as the rest of ScriptGrain's brand writing series and set against nine other sectors. The sector report behind SGR-010 and SGR-011. Free data under CC BY 4.0.

Conducted 2026-10-05, published 2026-10-11.

Sample

10 UK charities from ScriptGrain's brand deep corpus (October 2026): blog posts, press releases, product announcements and information packs from their own sites, fetched on 4 and 5 October 2026, plus archived pages from 2015 to November 2022. 1,994 pages measured. (n = 1994)

Method

• Plan: the sector reports were set out in the series pre-registration (docs/research/brand-series-preregistration.md section 6) before any page was measured: the SGR-010 tables for the sector against all other sectors, and the SGR-011 measures within the sector, exploratory, with intervals.
• Text and measures: as in SGR-010. Each page's main content reduced to plain text and counted in code.
• Weighting: each brand's pages are averaged first, then brands, here and for the other nine sectors.
• Years: pages with a date from their metadata, structured data or URL, or an archive capture date. A year is shown for the sector when it has 10 or more pages.
• Before and after: SGR-011's rule, inside the sector: a brand channel with at least 10 dated pages from 2019 to November 2022 and 10 from 2024 on.

What was measured

• Words per sentence
• Sentence length variation
• Long words (7+ letters)
• "You" share of I, we, you
• "We" share of I, we, you
• Contractions per 1,000 words
• Questions per 1,000 words
• Exclamation marks per 1,000 words
• Em dashes per 1,000 words
• AI-associated words per 1,000
• AI-era marketing words per 1,000
• Words per page

Findings

• Fewer AI-era marketing words than other sectors: 0.28 against 0.96 (−71%). Charities ranked lowest of 10 sectors. The middle half of the sector's brands ran from 0.16 to 0.40.
• Fewer AI-associated words than other sectors: 0.56 against 1.21 (−54%). Charities ranked lowest of 10 sectors. The middle half of the sector's brands ran from 0.50 to 0.67.
• Fewer questions than other sectors: 2.0 against 3.2 (−35%). Charities ranked 2nd lowest of 10 sectors. The middle half of the sector's brands ran from 1.2 to 2.7.
• Less "you" than other sectors: 34% against 51% (−33%). Charities ranked 2nd lowest of 10 sectors. The middle half of the sector's brands ran from 26% to 46%.
• AI-associated words by year: Peak in 2021: 1.06 per 1,000 words. Across all 10 sectors the peak was 2024 (2.16). Charities ranked lowest of 10 sectors on these words over the whole period.
• Press releases against the same brands' blogs: Longer sentences (+3.5 words, 95% interval 2.2 to 5.0); less "you" (−20.2 points, 95% interval −27.0 to −12.3). Eight charities had at least five of each. Positive numbers mean the press releases had more.
• Before and after ChatGPT (exploratory): 0.54 to 0.67 AI-associated words per 1,000; seven brands. Seven of the 10 charities had at least 10 dated pages in one channel both before ChatGPT (2019 to November 2022) and from 2024. In that writing, AI-associated words went from 0.54 to 0.67 per 1,000 words (+0.13, 95% interval −0.22 to 0.49).

Limitations

• 10 brands. A sector average can move with one or two brands; the brand range column shows the spread.
• Pages are what each brand publishes on its own sites. Brands that publish little have few pages.
• Counts of surface habits only. They describe style, not quality.
• Descriptive. Differences from other sectors are reported without significance tests.

Read the full study

SGR-019: Has the em dash really taken over? Em dashes per 1,000 words in books, research papers and brand writing, 1800 to 2026

How often did published writing use the em dash from 1800 to 2026, did its use in brand writing, research abstracts and company homepages rise after ChatGPT beyond its earlier trend, and is today's rate high or low against 19th-century books?

A pre-registered count of em dashes per 1,000 words across 2,059,475 arXiv abstracts, 7,464 dated brand pages, 2,950 company homepages and 1,276 Project Gutenberg books. Brand writing: 2.72 in 2019 to November 2022, 2.28 since 2024. Nineteenth-century non-fiction books: 4.89.

Conducted 2026-10-13, published 2026-10-14.

Sample

Five collections held by ScriptGrain: 2,059,475 arXiv abstracts (2015 to 2026, 333,063,827 words), 15,941 blog posts and press releases from 116 UK brands (7,464 of them, from 100 brands, dated well enough for the time comparisons), 2,950 company homepages (2000 to 2026, SGR-004), 1,276 Project Gutenberg books (108,615,578 words) and 6,600 web pages from 2021 (SGR-007). (n = 2077765)

Method

• Pre-registration: the six hypotheses, the dash rules and the analysis were committed to the public repository before the full measurement (docs/research/sgr-019-preregistration.md). The script was tested on under 5% of the data; the test results are not reported.
• Counting: one regular-expression count per text, no language model. Words are counted after dash marks are replaced by spaces, so "word--word" is two words.
• Trend test: for each modern body of writing, the 2015 to November 2022 pooled rate by quarter is fitted with a weighted line and projected over 2024 onward; the hypothesis is supported if the lower end of the 95% interval of observed minus projected is above zero. December 2022 to December 2023 is left out of both sides.
• Comparison with books: the ratio of today's rate to the rate in 1800s non-fiction books (and science books for arXiv), with independent bootstrap intervals; supported if the upper end is below 1.
• Breadth: the share of brands (with at least five pages in each period) whose rate was higher from 2024 on than in 2019 to November 2022, with a Wilson interval.

What was measured

• Em dash marks per 1,000 words (primary and strict)
• Share of texts with at least one dash (arXiv, brand writing, homepages)
• Source, year or era, genre, sector

Findings

• H1 (not supported): brand writing since 2024 sits above its 2015 to 2022 trend: Observed 2.28, trend 2.69, difference −0.41 (95% interval −1.06 to 0.27). Per 1,000 words, 100 brands. The rule was that the lower end of the interval must be above zero.
• H2 (supported): arXiv abstracts since 2024 sit above their 2015 to 2022 trend: Observed 0.40, trend 0.32, difference +0.08 (95% interval 0.06 to 0.11). Per 1,000 words, 150 subject categories. arXiv stores abstracts as ASCII, so only authors who type three hyphens or a spaced pair register here.
• H3 (not supported): company homepages since 2024 sit above their 2015 to 2022 trend: Observed 0.91, trend 0.94, difference −0.03 (95% interval −0.62 to 0.54). Per 1,000 words, 158 sites. SGR-004's counts, which include U+2014 only and are therefore a lower bound.
• H4 (supported, supported): today's rate is below the 19th century's: Brand writing 2.28 against 4.89 in 1800s non-fiction (ratio 0.47, interval 0.39 to 0.57); arXiv 0.40 against 5.32 in 1800s science (ratio 0.08, interval 0.05 to 0.12). The rule was that the upper end of each ratio's interval must be below 1. Book rates include editors' and typesetters' dashes as well as the authors'.
• H5 (not supported): brand writing was already rising before ChatGPT: 2.88 in 2015 to 2018, 2.72 in 2019 to November 2022 (ratio 0.94, interval 0.79 to 1.17). Supported if the lower end of the ratio's interval is above 1.
• H6 (not supported): the rise is spread across brands: 28 of 52 brands higher (54%, interval 41% to 67%). Supported if at least 60% of eligible brands rose and the lower end of the Wilson interval is above 50%.

Limitations

• Dashes in books may be the editor's or the typesetter's, not the author's, and the Project Gutenberg plain-text versions show the dash as two hyphens whichever mark the print edition used. The book rates measure the printed or transcribed text, not an author's habit.
• arXiv stores abstracts as ASCII, so an author's typed em dash survives only as three hyphens or a spaced pair of hyphens. Dashes written without spaces are missed, and the arXiv rate is a lower bound that depends on how authors type.
• The homepage counts are SGR-004's published rows: U+2014 only, with no spaced en dash, and the page text no longer exists to recount. They are a lower bound and are not pooled with the other collections.
• The collections differ in kind, so the lines in the long-run chart are not one sample of English over time. Books are literary and non-fiction works of different eras; the modern collections are marketing copy, scientific abstracts and company homepages.
• Brand pages come from fixed collection rules, not a random draw from everything a brand published, and many have no recoverable date. Archived pages (before November 2022) and live pages (after) came by different routes.
• A rise after 2022 shows a change in how text looks in the ChatGPT era. It is not evidence that any text was written by AI, and a fall would not be evidence of the reverse.
• Word counts here ignore dashes; SGR-010 counted every run of non-space characters, so rates can differ in the second decimal place.

Read the full study

SGR-020: Did government writing change after ChatGPT? 55,350 GOV.UK press releases and news stories, 2015 to 2026

Within the same UK government departments, did press releases and news stories published after ChatGPT use more of the words and habits associated with AI writing than those published before it?

A pre-registered before-and-after study of 55,350 GOV.UK press releases and news stories from 2015 to September 2026. Each of 20 departments is compared with itself: 2019 to November 2022 against 2024 onwards. AI-associated words rose 22%; em dashes did not. Free dataset under CC BY 4.0, source credited under the Open Government Licence.

Conducted 2026-10-15, published 2026-10-16.

Sample

Every GOV.UK press release and news story first published from January 2015 to September 2026 that GOV.UK's Search API listed on 11 October 2026: 62,399 items listed, 55,350 measured after exclusions, from 20 departments and 293 other bodies. The confirmatory tests use the departments only; 36 department-and-type pairs qualified. A census, not a sample. (n = 55350)

Method

• Corpus: every item of document type press_release or news_story on GOV.UK, listed through the Search API by month and fetched through the Content API, at no more than four requests a second with an identified user agent. 62,399 items were listed; embassy stories (world_news_story) and every other document type are not included. Source: GOV.UK, Open Government Licence v3.0.
• Departments: each item's primary publishing organisation. Twenty departments are analysed, with renamed, merged and split departments joined into one lineage (for example the Foreign and Commonwealth Office and the Department for International Development with the FCDO). Agencies, regulators and other bodies are analysed separately as exploratory.
• Text measured: the body of each item reduced to plain text, with tables, contact details and notes to editors removed; the title is not measured. Items under 150 words, exact duplicates within a department, withdrawn items, translations and items where more than 40% of the words sit in lines of six words or fewer were excluded.
• Measures: computed in code by ScriptGrain's own published instruments, the same as in the brand series. No language model read or scored any item.
• Pre-registration: the hypotheses, eras, exclusions and tests were committed to the public repository before the full measurement (docs/research/sgr-020-preregistration.md). The run was done once, on the day before publication.
• Tests: within-department change in department-weighted means, with 95% intervals from 2,000 bootstrap resamples of departments (fixed seed). H2 compares the per-year rate of change from 2015-18 to 2019-22 with that from 2019-22 to 2024-26, for cells with all three eras. H6 repeats H1 using only 1 January to 4 July 2024. Analyses beyond H1 to H6 are labelled exploratory.

What was measured

• AI-associated words per 1,000 words (and each word list)
• AI sentence shapes per 1,000 words
• Em dashes per 1,000 words
• Average sentence length
• Contractions per 1,000 words
• "You" share

Findings

• H1 (supported): AI-associated words rose within the same departments: +0.30 (95% interval 0.13 to 0.49) per 1,000 words, +22%. From 1.37 to 1.67 per 1,000 words across 20 departments. By word list: AI-era marketing words +54% (interval clears zero); the quieter LLM-favoured words +10% (within chance); the 2024 tells such as "delve" −4% (within chance); the publicly mocked ChatGPT words +28% (within chance). Leaving out any one department, the change stays between +0.26 and +0.34.
• H2 (not supported): the change after ChatGPT was not clearly different from the trend before: 0.067 per 1,000 words a year after, against 0.018 a year from 2015-18 to 2019-22. For the 20 departments with pages in all three eras, the difference in yearly change (+0.049, 95% interval −0.007 to 0.096) includes zero. These words were already moving in books and the web before ChatGPT, as SGR-007 found.
• H3 (not supported): em dashes did not rise: 1.77 to 1.92 per 1,000 words; change +0.15 (95% interval −0.18 to 0.42). Em dashes and spaced en dashes are counted together. Despite the 'ChatGPT dash' reputation, these departments used dashes at a rate within chance of the earlier one.
• H4 (supported): AI sentence shapes rose: 0.42 to 0.53 per 1,000 words; change +0.11 (95% interval 0.05 to 0.19). 'Not X but Y' constructions, reflexive lists of three, chains of semicolons and stock phrases, per 1,000 words.
• H5 (supported): sentence length stayed within 1.5 words: 22.9 to 22.8 words; change −0.12 (95% interval −0.52 to 0.32). The 90% interval (−0.47 to 0.25) sits inside ±1.5 words.
• H6 (not supported): the Conservative months of 2024 alone do not show a clear rise: +0.13 (95% interval −0.13 to 0.36) per 1,000 words (1 January to 4 July 2024 against 2019 to 2022). 20 departments qualified. 1 January to 4 July 2024 is six months and includes the election campaign, so this study cannot say whether the change of government or the time since ChatGPT explains the shift.
• Robustness and where the change sits: H1 with TRANSITION counted as after +0.22 (95% interval 0.06 to 0.40); items of 300+ words +0.31 (95% interval 0.10 to 0.52). Using the date of last update instead of first publication: +0.30, 95% interval 0.11 to 0.50. Press releases +0.42, 95% interval 0.21 to 0.64; news stories +0.13, 95% interval −0.15 to 0.36. Labour months (from 5 July 2024) against 2019 to 2022: +0.32, 95% interval 0.13 to 0.52.

Limitations

• The Content API serves each item as it reads now. Corrected or edited old releases are measured as edited; the dating check using the last-update date shows how much that matters.
• Ministers' quotations, shared boilerplate and lists of facts sit inside many releases and are measured with the rest; the press office's own prose cannot be separated from them.
• Word lists count spellings. These words rose in human writing for decades before ChatGPT, and topics shifted too: government now writes more about AI itself.
• Renamed and merged departments are joined into lineages by judgement. The business, energy, trade and science lineage mixes departments that did different work, and some work moved between departments in 2023.
• The 2020 and 2021 pandemic communications are inside the before period and changed what departments wrote; departments, not items, are weighted equally so the volume does not dominate.
• The Conservative-months check covers six months including an election campaign, so it is a weak test of whether the change of government explains the shift.
• UK central government only, in English, and only press releases and news stories.

Read the full study

SGR-021: Which classic authors write most like an AI? 1,129 authors, 6,293 books, measured

Which human writers, long before any language model, already used the words and habits now called AI-isms, and how much? Which of those words lived in which authors and eras?

6,293 books by 1,129 English-language authors from Project Gutenberg, ranked by how often they use the word lists now called AI-isms, with intervals over their books and a check on era, genre and dialogue. Pre-registered; published as a free dataset under CC BY 4.0. It ranks human habits; it does not detect AI.

Conducted 2026-10-20, published 2026-10-21.

Sample

The set is plainly "authors in Project Gutenberg's public-domain English catalogue who meet the inclusion rule", not a curated canon. It is made of English-language books (catalogue snapshot of 4 October 2026) by authors that Wikidata records as English-language writers with Gutenberg IDs and known birth and death years. An author is in if their midlife year (birth plus 40) is between 1650 and 1949 and they have at least three eligible books of 15,000 words or more; up to eight books each, chosen in a seeded random order. 1,129 authors, 6,293 books and 459.4 million words: 3 authors from the 1600s, 16 from the 1700s, 660 from the 1800s and 450 from the 1900s by midlife year, 3,324 fiction books and 2,969 non-fiction. (n = 6293)

Method

• Pre-registration: the inclusion rule, word lists, hypotheses and outputs were committed to the public repository before any book was measured (docs/research/sgr-021-preregistration.md). The panel is whatever the rules produce, with no hand edits.
• Measurement: each whole book is scanned with the frozen word lists from SGR-007 by deterministic code; a book's value is its uses divided by its words. No language model reads or scores any text.
• Ranking: authors ordered by total uses over total words, with a 95% interval from 10,000 resamples of their books, and re-ranked without their top book, against peers of the same era and genre, and on narration only.
• H1 to H4: pre-registered tests with 95% bootstrap intervals over authors (10,000 resamples).

What was measured

• AI-ism uses per 1,000 words
• AI-ism uses per 1,000 words of narration
• Uses of each listed word
• Share of words in dialogue
• AI-tell pattern density per 1,000 words (secondary)

Findings

• The ranking: who used the lists most: Richard Cannon 1.916, Charles Sturt 1.376, George Bryce 1.339 uses per 1,000 words. The corpus rate is 0.227 and the middle author's is 0.196. The next in the top 10 are Harry de Windt, John Charles Van Dyke, Mungo Park, James Augustus St. John, Ármin Vámbéry, Matilda Betham-Edwards and Esther Singleton. The chart shows intervals over each author's books; 4 of the top 20 get more than half their uses from a single book. Most of Richard Cannon's rate is one listed word: "embark" and its forms, 1.826 of the 1.916 uses per 1,000 words; the books are regimental histories, where regiments embark for campaigns.
• Robust to one odd book: 9 of the top 10 stay in the top 10 without their top book. 15 of the top 20 by rate are also in the top 20 once each author's top book is left out, and 14 are in the top 20 against peers of their era and genre. The robust top 10 is Richard Cannon, Charles Sturt, George Bryce, Harry de Windt, John Charles Van Dyke, Mungo Park, James Augustus St. John, Ármin Vámbéry, Matilda Betham-Edwards and John Richard Green.
• H1 (supported): non-fiction against fiction: +0.072 uses per 1,000 words; 95% interval +0.055 to +0.089. Non-fiction 0.263, fiction 0.191. Genre is the catalogue's label, and many authors wrote both.
• H2 (supported, a difference): 1900s against 1800s authors: −0.036 per 1,000 words; 95% interval −0.052 to −0.019. Mean author rate 0.206 for 450 authors whose midlife year was in the 1900s and 0.242 for 660 in the 1800s. We predicted no direction. Era and genre are both in the peers ranking, whose top 10 is Richard Cannon, Charles Sturt, George Bryce, Harry de Windt, John Charles Van Dyke, Mungo Park, James Augustus St. John, Ármin Vámbéry, John Bunyan and Matilda Betham-Edwards. Where an era has few authors the peers baseline is thin (the 1600s have 3), so John Bunyan (rank 737 by rate, ratio 3.76, interval 0.41 to 8.31) enters that top 10 with a wide interval.
• H3 (supported): not a dialogue artefact: Spearman 0.97; 95% interval 0.96 to 0.97. The bar was 0.9. 16 of the top 20 are also in the narration-only top 20. Exploratory: the author rate correlates with dialogue share at Spearman −0.19.
• H4 (supported): a stable habit of the author: ICC 0.43; 95% interval 0.33 to 0.54. The share of the variation in book-level rates that is between authors rather than between an author's own books. The bar was a lower bound above 0.25.
• Where the listed words lived: 38 listed words, each with a top author and an earliest strong home. "delve": 419 uses, 0.001 per 1,000 words, used most by Ernest William Hornung (0.034); no decade at twice the corpus rate with enough uses (peak: 1930s, 1.99 times). "tapestry": 2,155 uses, 0.005 per 1,000 words, used most by Esther Singleton (0.272); earliest strong home: authors who turned 40 in the 1800s (2.08 times the corpus rate). "realm": 7,894 uses, 0.017 per 1,000 words, used most by John Richard Green (0.702); no decade at twice the corpus rate with enough uses (peak: 1810s, 1.62 times). "embark": 8,605 uses, 0.019 per 1,000 words, used most by Richard Cannon (1.826); earliest strong home: authors who turned 40 in the 1750s (2.85 times the corpus rate). "journey": 42,874 uses, 0.093 per 1,000 words, used most by George Bryce (1.161); no decade at twice the corpus rate with enough uses (peak: 1830s, 1.60 times). "landscape": 10,335 uses, 0.022 per 1,000 words, used most by John Charles Van Dyke (1.062); no decade at twice the corpus rate with enough uses (peak: 1800s, 1.24 times).

Limitations

• Project Gutenberg texts come from particular editions; spelling and wording can be modernised or edited. Only words are counted, not punctuation.
• Eras are by the year the author turned 40, not the book's first publication, because the catalogue gives e-text dates.
• Genre is the catalogue's subject label, and the corpus mixes novels, essays, histories, science and sermons; the peers ranking corrects for the label, not for subject matter.
• The word lists contain ordinary English words with older senses (realm, journey, landscape, harness); a count is a count of words, never evidence that a passage is machine-like.
• Up to eight books per author, chosen at random; an author's rate rests on those books, and the interval shows how far they disagree.
• Only authors that Wikidata and the catalogue let us match; the public-domain record under-represents women and writers outside Britain and America, so the ranking describes this corpus, not writers in general.
• Correlation only: the study cannot show that any author influenced any model, or that measurement can identify AI writing.

Read the full study

SGR-022: Do ministers' written answers sound different since ChatGPT? 42,082 parliamentary answers and 17,292 written statements, 2015 to 2026

Within the same UK government departments, do written parliamentary answers and written ministerial statements given after ChatGPT use more of the words and habits associated with AI writing than those given before it?

A pre-registered before-and-after study of 42,082 written answers to parliamentary questions and 17,292 written ministerial statements from 2015 to September 2026, drafted largely by civil servants. Each government department is compared with itself: 2019 to November 2022 against 2024 onwards. AI-associated words rose 26% in answers; em dashes did not. Free dataset under CC BY 4.0; source credited under the Open Parliament Licence.

Conducted 2026-10-22, published 2026-10-23.

Sample

Every written question answered between January 2015 and September 2026 that Parliament's Questions and Statements API listed (668,536 answered, non-withdrawn, non-holding answers), sampled with a fixed seed to 60,000 (up to 5,000 a year, at least 200 per department a year where the department has that many), plus every written ministerial statement. 42,082 answers and 17,292 statements were measured after exclusions, from 19 departments and other bodies. The confirmatory tests use the departments only; 35 department-and-house pairs qualified. (n = 59374)

Method

• Corpus: written questions were listed through the UK Parliament Written Questions and Statements API by week of tabling, at no more than three requests a second with an identified user agent; a fixed-seed sample was drawn by year and department and its full answer text fetched; every written statement was listed by month and fetched. Source: UK Parliament, Open Parliament Licence.
• Sampling: for each year, up to 5,000 answers, with at least 200 per department where available and the rest shared in proportion to size; within each year and department, answers ranked by a hash of the question ID and the seed, first ones taken. Departments' means are weighted by the inverse of the chance of selection.
• Departments: each answer's answering body. Nineteen departments are analysed, with renamed, merged and split departments joined into one lineage (for example the Department of Health with the Department of Health and Social Care). Other bodies are analysed separately as exploratory.
• Text measured: the answer (or statement) reduced to plain text, with tables removed; the question and title are not part of the answer's text. Texts under 50 words, holding answers, exact duplicates within a department and texts where more than 40% of the words sit in short lines or more than half are numerals were excluded.
• Measures: computed in code by ScriptGrain's own published instruments, the same as in the brand series and SGR-020. No language model read or scored any text.
• Pre-registration: the hypotheses, eras, sampling design, exclusions and tests were committed to the public repository before the full measurement (docs/research/sgr-022-preregistration.md). The run was done once, on the day before publication.
• Tests: within-department change in department-weighted means, with 95% intervals from 2,000 bootstrap resamples of departments (fixed seed). H2 compares the per-year rate of change from 2015-18 to 2019-22 with that from 2019-22 to 2024-26, for cells with all three eras. H6 repeats H1 using only 1 January to 4 July 2024. Analyses beyond H1 to H6 are labelled exploratory.

What was measured

• AI-associated words per 1,000 words (and each word list)
• AI sentence shapes per 1,000 words
• Em dashes per 1,000 words
• Average sentence length
• Contractions per 1,000 words
• "You" share

Findings

• H1 (supported): AI-associated words rose within the same departments' answers: +0.14 (95% interval 0.03 to 0.25) per 1,000 words, +26%. From 0.54 to 0.67 per 1,000 words across 18 departments. By word list: AI-era marketing words +77% (interval clears zero); the quieter LLM-favoured words +13% (interval clears zero); the 2024 tells such as "delve" −13% (within chance); the publicly mocked ChatGPT words +169% (from a small base of 0.013 per 1,000 words; interval clears zero). Leaving out any one department, the change stays between +0.11 and +0.16.
• H2 (not supported): the change after ChatGPT was not clearly different from the trend before: 0.030 per 1,000 words a year after, against 0.005 a year from 2015-18 to 2019-22. For the 18 departments with answers in all three eras, the difference in yearly change (+0.024, 95% interval −0.027 to 0.069) includes zero. These words were already moving in books and the web before ChatGPT, as SGR-007 found.
• H3 (not supported): em dashes did not rise: 0.54 to 0.47 per 1,000 words; change −0.06 (95% interval −0.15 to 0.02). Em dashes and spaced en dashes are counted together. Despite the 'ChatGPT dash' reputation, these departments used dashes at a rate within chance of the earlier one.
• H4 (not supported): AI sentence shapes did not rise: 0.59 to 0.67 per 1,000 words; change +0.08 (95% interval −0.03 to 0.19). 'Not X but Y' constructions, reflexive lists of three, chains of semicolons and stock phrases, per 1,000 words.
• H5 (supported): sentence length stayed within 1.5 words: 25.0 to 24.9 words; change −0.13 (95% interval −0.48 to 0.20). The 90% interval (−0.41 to 0.15) sits inside ±1.5 words.
• H6 (not supported): the Conservative months of 2024 alone do not show a clear rise: +0.11 (95% interval −0.09 to 0.34) per 1,000 words (1 January to 4 July 2024 against 2019 to 2022). 18 departments qualified. Parliament was dissolved from 30 May to 4 July 2024, and this window holds only about five months of answers. This study cannot say whether the change of government or the time since ChatGPT explains the shift.
• Written ministerial statements (same six tests, secondary): AI-associated words +0.27 (95% interval 0.13 to 0.41) per 1,000 words (16 departments). Em dashes +0.33 (95% interval 0.13 to 0.56); AI sentence shapes +0.02 (95% interval −0.14 to 0.16); average sentence length −0.13 (95% interval −0.56 to 0.36). Statements are a census, fewer in number than answers, so intervals are wider.
• The questions MPs and peers write (comparison series): AI-associated words +0.04 (95% interval −0.06 to 0.14) per 1,000 words. Answers' change minus questions' change in the same departments: +0.10, 95% interval −0.04 to 0.24 (exploratory). Questions are written by MPs, peers and their staff, not by civil servants, so they show what a change in the language of Parliament itself would look like.
• Robustness: H1 with the transition year counted as after +0.12 (95% interval 0.02 to 0.22); answers of 150+ words +0.10 (95% interval −0.16 to 0.35). Unweighted means +0.13, 95% interval 0.02 to 0.25; pooled counts over words +0.17, 95% interval 0.02 to 0.32; without corrected answers +0.14, 95% interval 0.02 to 0.25; Commons +0.23, 95% interval 0.13 to 0.32; Lords +0.03, 95% interval −0.15 to 0.20. Labour months (from 5 July 2024) against 2019 to 2022: +0.15, 95% interval 0.03 to 0.28.

Limitations

• Written answers are drafted by civil servants, approved and given by a minister, and often reuse earlier answers, templates and shared lines; a numbers-only measure cannot separate the official's drafting from the minister's edit.
• Answers are a sample, not a census: a fixed-seed draw of up to 5,000 a year, weighted back to each department's output. Weighting corrects the design but not the exclusions that follow (short answers, duplicates), which fall unevenly across years.
• Many answers are very short, point to an earlier reply, or are mostly a list of figures. Texts under 50 words and mostly-numeric texts are left out, so the study describes substantive prose answers, not every answer.
• Word lists count spellings. These words rose in human writing for decades before ChatGPT, and topics shifted too: Parliament now asks about AI itself.
• Departments are joined into lineages by judgement. The business, energy, trade and science lineage mixes departments that did different work, and some work moved between departments in 2023. HMRC questions are answered through the Treasury and are counted there.
• Parliament was dissolved before the 2017, 2019 and 2024 general elections and prorogued between sessions (twice in 2019); no answers are given in those gaps, so months are not equally represented.
• The Conservative-months check covers about five months of answers (dissolution from 30 May 2024), so it is a weak test of whether the change of government explains the shift.
• Written questions and statements of the UK Parliament only, in English; oral questions and debates are not covered.

Read the full study

SGR-023: What reading age do UK brands write for? Readability of 115 brands' blogs and press releases, 2015 to 2026

What Flesch-Kincaid grade and UK reading age do UK brands' blog posts and press releases demand, how do sectors differ, are press releases harder to read than the same brand's blog posts, and did readability change between 2015 and 2026, including after the FCA's Consumer Duty came into force?

A pre-registered readability measurement of 15,540 blog posts, press releases and announcements from 115 UK brands in 10 sectors. The median brand's writing scores Flesch-Kincaid grade 10.7, a reading age of 15.7; supermarkets are hardest (18.4) and food and drink easiest (14.5).

Conducted 2026-10-27, published 2026-10-28.

Sample

15,540 blog posts, press releases and product announcements from 115 UK brands in 10 sectors, taken from the brand deep corpus v1.4: 9,893 blog, 4,914 press, 733 announcement pages. 4,196 are archived originals captured between 2015 and November 2022 and 11,344 are live pages fetched on 4 and 5 October 2026. (n = 15540)

Method

• Pre-registration: the formulas, the syllable counter, the text rules, three hypotheses and the analysis were committed to the public repository before the full measurement (docs/research/sgr-023-preregistration.md). The script was tested on under 5% of the data; the test results are not reported.
• Text: each page's main text turned into plain text as in SGR-010 to SGR-018; only lines ending in a full stop, question mark or exclamation mark are scored. Numbers and web addresses are not words; hyphenated words count as separate words.
• Syllables: fixed spelling rules, validated against the CMU Pronouncing Dictionary before the run, with the dictionary count used as a labelled sensitivity.
• Press against blog: for each brand with at least 10 press releases and 10 blog posts, the average grade of its press releases minus that of its blog posts; supported if the lower end of the 95% interval of the mean over brands is above zero.
• Change over time: for each brand and channel with at least 10 pages in 2015 to 2018 and 10 in 2024 onward, the change in average grade; supported (harder) if the lower end of the interval of the mean over brands is above zero, reversed (easier) if the upper end is below zero.
• Consumer Duty: the change in average grade for banks and insurers from before July 2022 (the month the FCA published the final rules) to after July 2023 (the Duty in force), minus the same change in all other sectors; supported if the upper end of the interval is below zero.

What was measured

• Flesch-Kincaid grade, UK reading age and Flesch Reading Ease
• Average sentence length and syllables per word
• Share of words of seven or more letters and of three or more syllables
• Share of sentences longer than 25 words
• Sector, channel, source, year

Findings

• H1 (supported): press releases are harder to read than the same brand's blog posts: +2.76 grades (95% interval 2.26 to 3.23), 48 of 53 brands harder in press. Press release average minus blog average, Flesch-Kincaid grade, brands with at least 10 pages in each. Of the difference, +1.67 grades came from longer sentences and +1.10 from longer words. Supported if the lower end of the interval is above zero.
• H2 (supported): writing became harder to read between 2015 to 2018 and 2024 onward: +1.38 grades (95% interval 0.85 to 1.95), from 9.77 to 11.39. Within brand and channel, 26 pairs with at least 10 pages in each period. The prediction was that writing became harder, so the rule was that the lower end of the interval must be above zero; a fall whose interval is below zero is reported as reversed. A pre-registered check on how the pages were collected matters here: for 36 brands with both live and archived pages from 2019 to November 2022, the live pages scored +0.96 grades against the archived ones from the same years (95% interval 0.62 to 1.29). Every page from 2024 on is a live page and most dated pages from 2015 to 2018 are archived copies, so part of the change over time may come from how the pages were collected rather than from the writing.
• H3 (not supported): banks and insurers became easier to read relative to other sectors after the Consumer Duty: Difference in change −0.34 grades (95% interval −1.55 to 1.00). 8 regulated brands changed +0.37 grades and 18 other brands +0.71. The regulated group did not clearly differ from the rest. Supported if the upper end of the interval is below zero.

Limitations

• Readability formulas measure sentence length and word length. They do not measure comprehension, layout, vocabulary a reader already knows, or whether the content suits its audience. A technical page for specialists can score as hard and still serve its readers.
• Reading age is the Flesch-Kincaid grade plus five, a convention for translating a US school grade, not a measured age of any reader.
• Syllables are counted by rules and are about nine in ten right per dictionary word, higher on common words. Names, acronyms and brand terms are approximate; the dictionary sensitivity shows how much that moves the results.
• Only running prose is scored. Pages whose content is mostly headings, lists and short labels are scored on what is left, and their scores describe that remainder.
• Pages were sampled by fixed rules from each brand's own sites, not at random from everything a brand published, and many pages carry no recoverable date. Archived pages (to November 2022) and live pages (after) came by different routes; the source check in the repository compares them.
• The Consumer Duty comparison cannot separate the Duty from anything else that affected banks and insurers differently over the same years. Some brands in other sectors, such as supermarkets, run FCA-regulated arms; one check removes them. The Duty's own wording concerns communications to retail customers, which may not be the blog and press pages measured here.
• UK consumer brands only, in the 10 sectors the earlier studies chose.

Read the full study