Could anyone have written this?
The Anyone Test
Paste in something you have published. You will get a score, the phrases your competitors also use, the sentences that are genuinely yours quoted back, and one thing to fix.
It will not rewrite anything for you. A tool that produced polished copy on demand would be the very thing smarticulate exists to object to, and there are already plenty of those.
It is not an AI detector either, and it is worth saying plainly that nothing which claims to be one actually works. This does not know, and does not care, whether you typed the words or prompted ChatGPT for them. Neither does the person reading them. The only question that costs you money is whether they could have come from anybody.
Nothing you paste leaves your browser. There is no server call. The scoring code came down with this page and runs on your own machine, so an unpublished draft, a board paper or a client proposal is safe to put in the box. Why that is architecture rather than a promise.
What it does and does not do
Four numbers, honestly labelled
Twenty-two things are measured. These four carry all 100 marks between them, and each one earned its place by separating six real businesses from a control written deliberately to look competent and mean nothing.
Own words. How much of this came out of the category phrasebook. The single most reliable signal there is, and the one that costs most sites their marks. It is also the one ChatGPT, Claude, Copilot and Gemini all push hardest in the wrong direction, because the phrasebook is exactly what they have read most of.
Rhythm. Whether your sentences vary. Writing that clusters around one length reads as wallpaper. Across the businesses measured, genuine voices scored between 7.7 and 9.7 on sentence-length variance. The generic control scored 4.3, and the two ranges never once overlapped.
Landing. Whether anything is short enough to hit. Real voices ran between 15% and 29% of sentences at five words or fewer. The control managed 0%. Roughly one in six is the target, placed where the argument turns.
Evidence. Numbers, results, named specifics. Of three unnamed sites measured while building this, two contained no figure of any kind, and one of those two leads with the word evidence in its headline. That is a claim arguing against itself.
Eighteen further things are measured, shown to you, and never scored. Several of them are exactly what you would expect a test like this to count, and they were disqualified for flattering the fake: the generic control used more contractions than two of the three genuine voices, and its sentences were longer than all of them. A measure that flatters a fake is not a test.
Below twenty-five sentences, rhythm and landing get noisy: a well-written 120-word passage can legitimately contain no five-word sentence at all. Where that applies, the tool says so rather than pretending. Own words and evidence are densities, and they hold at any length.
All twenty-two are set out in how the score works, along with where the scale came from and the four things this number is not.
Privacy, plainly
Architecture, not a promise
The difference matters. A privacy policy can be rewritten quietly and you would never know it happened. Architecture cannot be rewritten quietly: there is no endpoint here to send anything to, so there is nothing to change its mind about later.
The scoring code comes down with the page and runs on your own machine. Nothing is transmitted. Nothing is stored. Nobody sees it. Close the tab and it is gone. There is no account to create either, and no email address to hand over first.
People paste things into tools like this that they really should not. It happens constantly. The safest way to hold confidential text is to never receive it.