Copy that reads as AI-written, and what Docket can tell you

By Matt Kerr·Published 15 September 2026

You drafted some pages with a language model, and now you are wondering whether it shows and whether that costs you anything. Docket has a check for it — content.ai_slop, "Generic AI-sounding copy" — and the first honest thing to say is what it cannot do.

It cannot tell you whether a machine wrote your page

Nothing in the check looks at authorship, and there is no signal in a web page that would let it. What it does is match a short list of English stock phrases against your body copy. That is a test of register, not of origin. A human writer with a fondness for stock phrasing fails it. A model draft that somebody edited properly passes it, and should.

So read a finding from this check as "these pages sound like everyone else's pages", never as "these pages were generated". The second claim would be unfalsifiable, and a tool that makes unfalsifiable claims about your writing has stopped being useful.

What it actually computes

For every indexable page, Docket lowercases the body text and counts how many phrases from a fixed list appear in it. A page is reported only if it carries two distinct phrases from that list. One is never enough, and the reason is the most interesting thing about this check — see below.

Where any pages qualify, the check raises a single finding at medium severity covering all of them, titled that those pages read as generic AI-generated copy. The detail line quotes the first phrase it matched and the URL it matched on, so the finding carries its own evidence instead of asking you to take it on trust. It is marked as a large piece of work, which is honest: rewriting a page is not a settings change.

The fix text is a single sentence, and it is the part worth reading twice:

Rewrite in the business's own voice with specifics only you have: real prices, real timelines, named neighbourhoods, actual customer situations.

That is not a style note. Stock phrasing is a symptom; the disease is that the page contains nothing a competitor could not have written. Prices, timelines, place names and real customer situations are the things a model cannot invent for you, because it does not know them. Supply those and the stock phrases tend to fall out on their own, because there is finally something for the sentence to be about.

Why two phrases, and the entry that was removed

The threshold is the check's defence against convicting ordinary writing, and it was very nearly not enough.

The list once contained an ordinary English connective that predates the language models by a very long way. It was measured against human-written pages from institutional sites — a church, a credit union and a trade union, none of them named here — and it turned up on every one of them. Every other marker in the list turned up on none of them. That is not a tell; it is a joint in the language.

Because the check fires on a pair, that one entry was enough to turn a single genuine cliché into a conviction. On a large tutorial site, one page was reported on the strength of a real cliché in the article and the ordinary idiom in a reader's comment underneath it — the page was judged partly on somebody else's writing. The entry is gone, and the test that removed it also asserts that no remaining marker fires on plainly-written institutional English, so the next addition has to clear the same bar.

The comment half of that problem is still open and is worth knowing about if you run a blog with comments enabled. The text Docket reads includes the comment thread, and no check currently knows where the article stops and the readers start.

The list is short on purpose, and this site cannot print it

A long phrase list produces false positives on perfectly good human writing, so the list is deliberately small and every entry has to earn its place. The entries that survive are the ones that read as machine-written to almost anybody: the stock opener, the stock transition, the stock superlative.

There is a small joke in the fact that this page does not quote them. This site runs its own copy gate over every page it builds, and that gate bans nearly all of the same phrases as visible text — including inside a code block. A page printing the list in full would fail the build that published it. If you want the entries, they are legible in the check's source rather than on a marketing page, which is where a claim about your writing ought to be auditable from.

What it does not read

Each of these exclusions exists because the naive version was wrong about somebody:

• Code samples. The text these phrase checks read has pre, code, samp and kbd stripped out of it. Docket once told a popular CSS framework that its install guide contained placeholder text; the text was Hello world! inside the code block the guide exists to show. Every documentation site, tutorial and programming blog was exposed to the same reading.
• Pages you have excluded. The check reads indexable pages only, so anything carrying a noindex directive is not judged.
• Sites that are not in English. The check is phrase matching against English, so on a site whose content language is anything else it stands down entirely and the report says that it did. The alternative — running an English phrase list over Japanese and reporting every absence as a fact — is how a tool ends up telling a confectioner it has no call to action on pages that all have one.
• Everything else about quality. This check has no opinion on whether your page is accurate, useful, or worth reading. It matches phrases.

Docket's per-check reasoning, including this one, is listed with every other check it runs on what Docket checks.

Does any of it actually cost you traffic?

Google's published position is narrower than the panic around it, and worth reading directly. Its guidance on creating helpful, reliable, people-first content says that using automation, AI generation included, to produce content for the primary purpose of manipulating rankings violates its spam policies. Its spam policies define scaled content abuse as many pages generated primarily to manipulate rankings rather than help users, and say that such pages provide little value to users no matter how they were created.

Read those two together and the test is purpose, not production method. A page drafted by a model, edited by someone who knows the business and genuinely useful to a reader is not what either document describes. A mass of near-identical pages spun to catch a keyword pattern is, whoever or whatever typed them.

The second cost is the one nobody publishes a policy about. Readers now pattern-match stock phrasing instantly, and a prospect who decides on your first paragraph that nobody at the company wrote this has already decided something about the company. That is not a ranking problem and no crawler will ever measure it for you.

What to do when it fires

Open the pages the finding names and ask one question of each: what does this page contain that only we could have written? If the answer is nothing, the stock phrases are the least of it. Replace the generic paragraph with the specific one — what the job costs, how long it takes, which streets you cover, what the awkward customer situation was and how it was handled. Then read the result aloud. Stock phrasing is much easier to hear than to see.

If the same generic register runs across the whole site rather than a few drafted pages, the problem is a level up from copy. Docket's brand lane asks whether a person moving between your pages sees the same company each time, which is a question no crawler asks and a different question from this one.

And if the pages in question exist to answer a question that an AI assistant now answers directly, rewriting them in your own voice may not be the whole job. Which of your pages an AI answer can replace, and what defends one, is measured separately.

How much to trust this finding

Medium severity is the right weight for it. This is a heuristic over a small phrase list, it knows nothing about authorship, it can still be fooled by a comment thread, and it will miss machine-written copy that avoids the stock phrases entirely — which is most of it, and increasingly so. A clean result here is not a certificate that your writing sounds human. It only means these particular phrases are not doing it.

That asymmetry is worth carrying into every finding any audit tool shows you. The general version of the question is how to tell whether an audit tool is lying to you.

Download Docket

Common questions

Can Docket tell if my content was written by AI?

No, and neither can anything else reliably. The check matches a short list of English stock phrases against your body copy, which is a test of register rather than of authorship. A human writer fond of stock phrasing fails it; an edited model draft passes it.

What triggers the generic AI copy check?

A page has to carry two distinct phrases from the check's fixed list before it is reported. One is never enough. The finding is raised once at medium severity covering every page that matched, and its detail line quotes the phrase it found and the page it found it on.

Why does it need two phrases rather than one?

Because a single phrase convicts ordinary writing. One entry in the list was a common English connective that turned up on human-written pages across several institutional sites while no other marker turned up at all. It was removed, and the test that removed it now asserts that no remaining marker fires on plainly-written English.

Does AI-written content get penalised by Google?

Google's spam policies target purpose rather than production method: content generated primarily to manipulate rankings rather than help users, which they say provides little value no matter how it was created. An edited draft that genuinely helps a reader is not what that describes.

Does the check run on a site that is not in English?

No. It is English phrase matching, so outside English it stands down and the report says so, rather than finding nothing and reporting the absence as a fact about the site.

Does it read my code samples or my comment threads?

Code samples are stripped before matching, because a placeholder inside a code block is the subject of a tutorial rather than unfinished copy. Comment threads are not yet separated from the article, so a page can in principle be matched partly on a reader's words. That gap is known and recorded.

What should I do if my pages are flagged?

The check's own advice is to rewrite in the business's own voice with specifics only you have: real prices, real timelines, named neighbourhoods, actual customer situations. Those are the things a model cannot invent for you, and supplying them usually displaces the stock phrasing by itself.