AI words to avoid: all 98 words and phrases that make copy sound generated, ranked by real use

Most lists of "AI words" are somebody's impressions. A few famous examples, a few more that sound similar, and no indication of which ones actually turn up on real websites or how often.

This one comes from measurement. SiteTell checks every page it crawls against 98 words and phrases, and below is every one of them, grouped by what it does in a sentence and ranked by the share of real sites carrying it. Of the 80 sites whose stored page text we rescored under the current ruleset, 54 had at least one: 68%.

Two things to know before reading the tables. The words people joke about are mostly not the common ones. And a word appearing on your site is a prompt to reread a sentence, never proof of who wrote it.

How to read these tables

Sites is the share of those 80 sites where the word appeared at least once, using stored text from each site's latest scan, rescored on 2026-09-28 under ruleset 2026-09-16. Share of sites is steadier than raw counts, because one large site repeating a phrase across 400 pages can outweigh everyone else put together. That happens here: a single site supplies 81% of one bucket's hits for landscape.

An asterisk marks an ordinary English word, one where a share of hits are innocent. Software has between services, geography has a , and a test suite is a technical term. Those are flagged anyway, and the asterisk tells you to judge each one in context.

Severity is how much a hit costs a page: low, medium or high, worth two, four or seven points, with repeats beyond the third on a page damped. Plain swap is the replacement the product offers where there is a clean one. Where it says "rewritten in context", the word has no reliable one-for-one substitute and a fix has to read the sentence.

Rows at 0% are real rules that fired on none of these sites. They stay in, because a list trimmed to what happened to appear would quietly misstate what is being checked.

The famous words are not the common ones

The word everyone names first, delve, appears on 13% of these sites. tapestry manages 4%, and testament 8%.

Compare the top of the verb table. unlock is on 38% of sites, enhance on 35%, leverage on 33%. They are dull, which is exactly why they survive editing: nobody deleting chatbot clichés goes looking for them.

If you only have time to check a handful, check those three, plus landscape (31%) and robust (29%).

Verbs that inflate an ordinary action

The largest family and the most common. Each of these takes something plain (using a thing, improving it, making it simpler) and dresses it as an event. The fix is almost always the plain verb, followed by the specific result: what got faster, by how much, for whom.

Several of these match every verb form. The rule counted 366 base-form hits, 264 of the third-person form, 47 past tense and 27 gerunds. For , the gerund is nearly as common as the base form, at 385 against 419, because it so often opens a feature description.

Adjectives that praise without evidence

These describe a quality the reader has no way to check. A product called , a platform called : each asks to be believed rather than showing anything. Where the swap column offers a plainer adjective, treat it as a stopgap. The real fix replaces the adjective with the fact that would have justified it.

Metaphors and borrowed nouns

Words that turn a market into a , a combination into a , a field into a . One metaphor is style. Metaphors doing the job of a noun that already exists read as padding, and "market", "mix" and "area" are usually right there.

Openers, transitions and preambles

Phrases that fill the start of a sentence while the point waits. leads this group at 18% of sites. Most can simply be deleted, and the sentence starts where it should have. SiteTell also runs a separate structural check for stock sentence openers, so a few of these can trip two rules at once.

Hedges and puffed-up phrases

Multi-word constructions that promise significance: a thing that plays a role, a result that stands as a monument to something. These are rarer than single words, since each needs several words in exactly the right order, but when one appears it is rarely innocent, which is why most carry high severity.

Claims with no source attached

Three phrases that cite authority without naming it. They appear on few sites, but each carries high severity, and the product's advice is the same for all three: name the source, or cut the claim. A reader who wants to check has nowhere to go, and a reader who does not will skim past.

Chatbot leftovers

Phrases a chat assistant says to the person it is talking to, which make no sense on a published web page. They almost never appear, and most show zero hits here. When one does, it is close to certain that generated text went live without anyone reading it, which is why the whole group is rated high.

Phrases that stack

Some rules overlap by design. A sentence containing trips both the word rule and the phrase rule; the same holds for , , the two phrases built on , and . One phrase can therefore cost a page up to fourteen points on its own.

We left the overlap in deliberately. The longer phrase is a stronger signal than the word inside it, and charging for both reflects that.

Should you delete every word on this list?

No. A find-and-replace pass through this table will produce copy that avoids 98 words and still reads as generated, because the patterns that appear on the most sites are structural. Closing paragraphs that repeat the page, uniform sentence lengths and runs of three-item lists each appear on more sites than any single word here. Our walk through the whole ruleset covers those.

What the list is good for is finding sentences to reread. Most of the time, the flagged word is standing in for a specific claim the writer never made. Put the claim in, and the word usually goes with it.