No single word has come to stand for machine-written text quite like . It is the example people reach for first, the one that gets joked about, and in some circles the one that gets an essay rejected on sight.
So here is a number that may surprise anyone who has been scrubbing it from their drafts. Across the 80 sites whose stored page text we rescored under SiteTell's current ruleset, 10 used it. That is 13%. Plainer, blander words turn up far more often: on 38% of those sites, on 35%, on 33%.
This page argues two things at once. The word does belong on our ruleset, and we flag it. But it is a weak thing to judge a page by, and the fame it has picked up makes it a worse guide than the quieter words beside it.
- Rule
- Type
- vocabulary · AI cliché
- Severity
- medium
- Share of sites
- 13% (10 of 80)
- Instances
- 99 on 73 pages
- Per 1,000 words
- 0.012
- Forms matched
Does using “” mean AI wrote it?
On its own, no. The word is centuries old, it has a perfectly good literal sense (digging, reaching into a bag, searching through records), and plenty of careful writers use its figurative sense without a model anywhere near them.
What changed is frequency. A 2024 study of 15 million biomedical abstracts tracked words whose use jumped far beyond their historical trend after chatbots became widely available, and used that excess vocabulary to estimate that at least 13.5% of 2024 abstracts had been processed with a language model, reaching 40% in some groups. Words like this one became famous because their rise was so steep that it could be measured from outside.
That is a statistical signal about large bodies of text. It says very little about one sentence, which is the level at which people usually apply it.
How often does it actually appear on websites?
Rarely, and unevenly. The 99 flagged instances in our rescored data sit on 73 pages across those 10 sites, which works out to 0.012 per 1,000 words, one of the lower rates among the words anyone would call famous.
The unevenness matters more. SiteTell runs a second rule for the phrase , and 60 of those 99 hits are that phrase, found on just 3 sites. A handful of sites using it heavily, mostly in blog introductions, accounts for most of what exists. Everywhere else it is scarce.
The inflections tell a similar story. The rule matches every verb form, and the base form accounts for 85 hits, against 7 for the third-person form, 5 for the gerund and 2 for the past tense.
Why does SiteTell still flag it?
Because when it does appear, it tends to arrive with company. Pages carrying it are about 18 times likelier than an average page to also carry , and about 8 times likelier to carry . Both pairings span at least five separate sites. The same pages are four times likelier to trip the check for low vocabulary diversity, which catches prose circling a single point.
That clustering is the honest case for the rule. A reader who notices the word is often noticing a register: a whole paragraph written to sound like content, of which this word is the most recognisable piece. The flag points you at that paragraph, and the paragraph is usually what needs the work.
It also costs more than it looks. The phrase and the bare word are separate rules, and a sentence containing the phrase trips both, at medium and high severity. A single costs a page eleven points before any other flag is counted.
What it looks like, and a better sentence
These examples were written for this page rather than taken from a real site, since a quoted sentence could identify one. On every build of this site, each flagged version must trip the rule, and each rewrite must pass every rule SiteTell can check on a single passage.
The most common home for it is the blog introduction, announcing what the post is about to do. The fix is to skip the announcement and start doing it.
Both rewrites replace a verb of vague investigation with the specific action. "Click a bar to see the rows" and "interview two customers" can each be checked. Digging deeper cannot.
When is it a false positive?
More often than for most words on the list, and our own scan results show it. Among the pages flagged for it were a piece of fiction using the word in its older sense and a page listing it as a word to avoid, which is exactly the problem this article would have if its examples were plain text.
The rule matches on the word boundary with no sense of meaning, so a mining company, an archive or a novelist will see flags. Each one quotes its sentence, so dismissing them takes seconds. What we would argue against is the opposite mistake: treating any single word, this one included, as proof of anything. A page with it and nothing else wrong is almost certainly fine. A page without it can still read as generated from top to bottom.
If not this word, what should you watch for?
The words that actually saturate marketing copy are the dull ones. Measured by share of sites, the leaders in our data are verbs that inflate an ordinary action and nouns borrowed as metaphors, and their full list, with how often each appears and a plain alternative, is on our page of AI words to avoid.
Past the word level, the patterns that appear on the most sites are structural: closing paragraphs that repeat the page, the em dash used as a crutch, lists of three on repeat. None of them is famous, and each one is on well over half the sites we measured.