Landing Copy Grader

July 25, 2026 · View on GitHub

DOI

Does your landing page hero copy read as AI-generated? Paste your headline, subheadline and CTA, get a 0–100 score plus the exact lines to fix.

No LLM. No backend. No signup. One HTML file, ~15 KB, runs entirely in your browser — open it from file:// on a plane if you want. Your text never leaves the page.

▶ Try it in the browser · hosted version · or download grader.html and double-click it.

node scripts/score-page.js https://stripe.com
# headline: Financial infrastructure to grow your revenue. …
# subhead : Flexible solutions for every business model.
# cta     : Get started
# score   : 61/100 — Decent, but softening in places.  [filler,weakcta,nonum,longhl]

The repo also ships the dataset that came out of it: the hero copy of 239 real landing pages, extracted and scored with the grader in this repo, under the rules of that date. The headline finding — 195 of the 239 pages (82%) contain no number at all in their hero. Median score 79. node scripts/verify-dataset.js re-scores all 239 rows offline and fails if a single one disagrees.

Screenshot of the grader scoring a hero 100/100 with per-dimension bars

Why deterministic instead of an LLM?

The obvious way to build a "does this sound AI-written?" tool is to ask an LLM. For this specific job, a fixed rule set is the better engineering choice:

Deterministic (this tool)LLM call
Same input → same scoreAlwaysNo — non-deterministic, drifts across model versions
LatencyInstant, localNetwork round-trip
Cost / keyFree, no API keyPer-call cost, key management
PrivacyText never leaves the browserText sent to a third party
ExplainableYou can read the exact rule that firedOpaque; may hallucinate its reasoning
OfflineYesNo

There's an irony worth stating plainly: asking one language model whether copy "sounds like a language model" is circular. The tells that make copy read as generated — hype verbs, zero specifics, filler nouns, weak CTAs — are surface, countable patterns. You don't need a 100B-parameter model to count them; you need a good list and a scoring rubric. That's what this is.

Honest limitations

This is a linter for copy, not a ground-truth AI detector. It scores stylistic tells, not authorship. Human copywriters write hype too, and a careful LLM can score 100. Treat the number as a fast heuristic — "here are the generic patterns in your text" — not a verdict on who or what wrote it. It's English-only and tuned for short hero copy (headline / subhead / CTA), not long-form.

The heuristics

Five weighted dimensions sum to 100. Every rule is a plain function over the three input strings — read the whole thing in grader.html (the scoring lives in one grade() function, ~40 lines).

1. Anti-hype — 25 pts. Starts at 25, subtracts penalties for the top "generated" tells:

  • hype words (revolutionize, unlock, seamless, leverage, supercharge, cutting-edge, game-changer, … ~40 terms) — −7 each
  • exclamation marks — −5 each
  • emoji — −4 each. Arrows (→, ➜) and check marks (✓) are not counted: a button arrow is typography, not an emoji.
  • shouting — −4 each. A word in caps counts only if it is 6 letters or longer, or one of 17 words people genuinely shout in a headline (FREE, NEW, SALE, …). SQL, MCP, CLI, API are acronyms, not shouting.

2. Specificity — 25 pts. A concrete number → 25. No number → 8, with a partial rescue (+9) if there's a proof-shaped word (%, x, hours, days, minutes, no, zero). One concrete figure is the single fastest believability lift.

A digit only counts as a claim if it isn't part of a name, a version, a year or a list index — Auth0, Mem0, n8n.io, Framer 3.0, B2C, © 2026 and a page quoting pt-4 are not making a quantified claim.

3. Clarity — 25 pts. Starts at 25, −6 per filler word (solutions, platform, powerful, amazing, experience, journey, ecosystem, all-in-one, … ~27 terms). These add length, not meaning.

4. Headline shape — 13 pts. Rewards a repeatable length. 3–10 words → 13; 11–12 → 9; >12 → 5; <3 → 7; empty → 0.

5. CTA — 12 pts. A specific action → 12. A generic CTA (submit, learn more, get started, sign up, click here, … ~14 terms) → 4. Empty → 0.

Word matching is whole-word and case-insensitive (regex-bounded, so learn inside learned doesn't false-positive).

Then a verdict band: ≥80 reads human & sharp · 60–79 decent · 40–59 somewhat generic · <40 reads AI-generated. You get up to six targeted fixes, each naming the exact count it found.

The three exclusions above are corrections, dated 2026-07-25, and each came from a measured false positive rather than a preference: 11 of the 44 pages credited with a number were name/version/year artefacts, arrows were costing pages 4 points for a button glyph, and every ALL-CAPS flag raised on a technical page was an acronym. Scoring a Postgres tool down for "SQL in capitals" discredits the tool that is selling rigour.

The dataset below predates all three and is scored with the previous rules. Its method column records which (static-fetch-regex-v1), and verify-dataset.js scores each row with the rule its own method names — re-scoring an archived row with today's rule would not be a verification, it would be a different dataset.

Worked examples

Reproducible — run them yourself:

CopyScoreVerdict
"Revolutionize your workflow with our seamless, cutting-edge platform / Unlock powerful solutions that transform your business / Learn more"32This reads AI-generated.
"Cut invoice time from 3 days to 20 minutes / Turn your spreadsheet into a client-ready invoice, no template hunting. / Start your first invoice"100Reads human & sharp.

The first triggers: cut the hype words, add a number, delete filler, rewrite the CTA.

Dataset: 239 real landing pages, scored

data/landing-pages-scores.csv — the hero copy of 239 well-known landing pages (YC companies, dev tools, SaaS, AI products), extracted on 2026-07-24 and scored with static-fetch-regex-v1, the rule set of that date. Live table: leaderboard.

One row per page, with the extracted text included — url, domain, score, flags, headline, subhead, cta, hero_chars, extracted_at, method — so every score is reproducible offline, without refetching anyone's site:

node scripts/verify-dataset.js     # re-scores all 239 rows, exits 1 on any disagreement
node scripts/score-page.js https://stripe.com    # or re-extract a page live

⚠️ Reproducing a row means reproducing it under static-fetch-regex-v1 — grade(headline, subhead, cta, 'static-fetch-regex-v1'), which is what verify-dataset.js does automatically by reading each row's method column. The browser tool applies today's rules and takes no method argument, so pasting a row into it can legitimately give a different score. That is the rule change below doing its job, not a contradiction — but you should know which of the two you are running.

What the corpus says, recomputed by that script — these are the figures of the deposited dataset (static-fetch-regex-v1), which is the archived, citable object:

tellpages%
nonum — no digit anywhere in the hero19582%
filler — ≥1 filler word8234%
weakcta — CTA is a stock phrase3515%
caps — ALL-CAPS word3314%
hype — ≥1 hype word167%
shorthl — headline under 3 words135%
longhl — headline over 12 words94%
emoji73%
excl — exclamation mark73%

Scores: min 41 · median 79 · mean 80.1 · 19 at 100 · 31 below 70.

The most common tell is not the em-dash and not "delve" — it's the absence of a number. Four landing pages in five make a claim with no quantity attached to it.

Correction, 2026-07-25. An earlier version of this CSV stored only the first three flags per row. Rows with four or more tells silently lost one, so five of the nine frequencies above were published low — nonum in particular read 194 / 81% instead of 195 / 82%. The CSV now carries the full flag list and the source text, and verify-dataset.js recomputes the table from it, so the published summary can no longer drift from the data. If you saw 194 / 81% from us anywhere, 195 / 82% is the correct figure.

Extraction method (canonical, documented in scripts/score-page.js): plain GET (no JS execution), first <h1> (og:title/<title> fallback, repeated-phrase collapse), first <h2>/<p> after it (20–400 chars, cookie boilerplate skipped), first <a>/<button> after it (2–40 chars, skip-links/consent/login UI skipped). Pages rendering <200 chars of text without JS were rejected.

Honest caveats: the grader sees only what a no-JS fetch returns — the extracted hero can differ from what a human sees on a JS-rendered page. Excluded from the dataset: pages served in a non-English locale, pages behind bot protection (403/503), pages intercepted by the crawl network's filter, and 4 pages where extraction produced a non-hero fragment (airbnb.com, dev.to, checkout.com, substack.com). Scores judge the extracted hero copy against fixed heuristics — they are not a judgment of the product or the full page.

Use it

Project status

Young and small, stated plainly so you can judge it: first published 2026-07-24. The scoring engine and the extraction method are stable — they are the two things this repo is for, and verify-dataset.js pins both against the 239-page corpus, so a change that moves a score breaks the build rather than the data.

What is likely to change: the word lists (they are opinionated, and PRs adding or removing terms are the most useful contribution), the corpus (it can be re-extracted; the pages move), and language coverage (English only today).

Issues and pull requests are welcome — including "this rule is wrong, here's a counter-example." A counter-example against a deterministic scorer is a reproducible bug report, which is most of the reason for building it this way.

Cite this

Both objects are archived on Zenodo with a DOI, so the rubric and the corpus can be cited independently.

They are separate records because they are separate things: the scorer gets corrected, the archived corpus does not. GitHub's Cite this repository button reads CITATION.cff and hands you BibTeX or APA directly.

License

MIT — see LICENSE. Fork it, ship it, rip out the parts you don't like.

Built by the 1h Money Store. The hosted grader is a free tool; the store sells copy prompt packs and a landing-page template. No obligation — the grader stands on its own.

Same engine, other surfaces:

Other repos by the same org:

  • god-flight-recorder — Flight recorder of an autonomous AI org running a real business. All decisions on file.
  • claude-swarm-starter — Run your own org of Claude agents coordinated through plain files.
  • leverage-dev-rules — Cursor rules for solo founders shipping their own product.
  • studio-starter — Free single-file HTML landing page starter — editorial serif, no build step, MIT.