Repetition

Repetitive AI content is a symptom of generation, not a style choice.

A model does not remember what it already told you. It reconstructs the argument each time, which is why the same point arrives three times in slightly different clothes.

Run it here

Count the openers and the loops in your own draft.

Full verification report
Repetition counter

Words

0

Flags

0

Per 1,000

0.0

Reading

Nothing flagged

No term from this list appears in the text.

Paste a draft. The counter flags repeated openers, self-references, summary loops and circular qualifiers, then shows how often each returns. Counted in your browser. Nothing you paste leaves the page.

Repetition is a symptom, not a style

An editor cutting a page worries about saying too little. A model has the opposite problem: it keeps asserting the same idea because each sentence is generated from the text before it, and the safest next sentence is a restatement of the last.

That is why repetition in generated prose is rarely word-for-word. It is conceptual: the same claim in five registers, each one convincing in isolation and cumulatively exhausting.

The cost is not only length. Repetition signals to a reader that nothing is being added, and a reader who suspects that starts skimming. The page can be accurate and still lose the person reading it.

Treat repetition as a structural finding rather than a wording problem. Moving paragraphs is usually part of the fix; swapping synonyms never is.

The five kinds of loop

Most repeats fall into one of these shapes. Naming the shape tells you what the fix is, because each fails differently.

  • Lexical loops: the same distinctive word or phrase returning every other sentence, so the draft circles a term instead of advancing it.
  • Restated theses: the opening claim reappearing as a paragraph's conclusion, with no new evidence between the two.
  • Parallel paragraph shapes: each section built to the same recipe, so the third paragraph of every section can be predicted.
  • Recycled examples: one story carrying three different points because no second example was ever found.
  • Hedging recursion: a claim softened, then softened again in the next sentence, until the paragraph asserts nothing that could be checked.

Why a model repeats itself

Text generation has no plan to consult. Each step conditions on what came before, so a distinctive phrase in the context window becomes more likely to appear again, and a point already made looks like a natural thing to make again.

There is also a safety pressure. Restating the thesis is rarely wrong, while introducing a new claim risks contradiction. Average out enough safe choices and you get a page that says one thing eleven times.

Long outputs amplify it. Under a thousand words the loop is invisible; at four thousand the draft is visibly treading water, and the reader feels the stall before they can point to it.

This is why a longer draft needs an outline an editor wrote. The outline is the memory the model does not have.

The opener problem

Count the first word of each paragraph. If furthermore, additionally and moreover take turns, the page has a visible mechanism, and readers who skim start noticing the pattern instead of the content.

Openers are also where a model hides a missing transition. The real relationship between two paragraphs might be concession, cause or exception, and a generic connective papers over the fact that the writer never decided which.

Delete the opener and read the two paragraphs together. If they now collide, the section order is wrong, not the vocabulary. Move the paragraph rather than finding a better adverb.

A useful constraint: no two paragraphs in a row may begin with the same part of speech. It forces variety at the exact place a template would show.

Repetition across a site, not just a page

The same mechanism operates at the level of a set of pages. Twenty drafts from one brief share an opening move, a structure and a metaphor, so a reader who lands on three of them sees one page written three times.

Search engines resolve this by picking one version, which means the other nineteen are competing with each other for nothing. The work was done and the value is diluted.

The test is whether two pages could be swapped without a reader noticing. If so, one page is doing the job of both, and the second should be merged, redirected or given material of its own.

Differentiate by substance rather than by wording. A page that exists because the query is phrased differently is duplicate content with a new title, and it is usually obvious within a paragraph.

Measuring it instead of feeling it

Two counts catch most of it. Unique words as a share of total words falls as a draft repeats itself, and the gap between the two numbers is worth watching as a page grows.

The second is opener frequency: how many paragraphs start with the same handful of words. Over about a fifth of paragraphs is a pattern; over a third is a template nobody edited.

Read the first sentence of every paragraph in sequence. That list is your outline. If it repeats itself, the draft is repeating itself, and you can fix the outline without touching 90% of the text.

Neither number is a verdict. Both are cheap ways to find the two or three paragraphs that need attention, which is faster than rereading four thousand words hoping to notice.

Editing without stripping emphasis

Some repetition is rhetorical. A phrase repeated at the start and end of a section can frame it. A warning repeated after each step is insurance, not padding.

The distinction is function. Repetition that adds emphasis is deliberate, spaced and usually short. Repetition that fills space is adjacent, paraphrased and often a full paragraph long.

Keep the deliberate instances and make them identical, so the reader recognises the return. Cut the accidental ones, which are the ones that differ slightly each time.

When in doubt, ask what the reader loses if the second instance goes. If the answer is emphasis, keep one of them and cut the other. If the answer is nothing, cut it.

Templates that manufacture repetition

A section template with four fixed slots produces four paragraphs that behave identically. Multiply by ten sections and the page has one shape, repeated.

Fix the template rather than the page. Allow sections with two paragraphs and sections with six. Permit a table, a list, a single example. The variance should be written into the brief so it survives the next draft.

The same applies to the models you reuse. A prompt that requests background, key points and a conclusion will produce a loop, because key points and conclusion are usually the same sentences wearing different hats.

Ask instead for the new thing each section contributes, and let the writer decide the shape that delivers it. Structure should follow the argument, not a fixed count of headings.

A working pass

Start with the outline: list every paragraph's first sentence. Merge or cut the ones making the same point. Most drafts lose a tenth of their length here without losing a claim.

Then handle the openers: delete each connective and check whether the logic survives. Where it does not, name the relationship in the sentence itself.

Then the examples: if one story is carrying three points, find two more stories or accept one point. An example stretched past its usefulness reads as repetition to anyone who follows it.

Finally check the close. If the last paragraph restates the introduction, cut one of them. Pages that begin and end in the same place tell the reader nothing happened in between, which is a repetition of a different kind.

When repetition is the point

Legal notices, safety instructions and definitions repeat on purpose, because a reader may arrive at any paragraph first. Do not edit those for variety; preserve the exact wording so the meaning cannot drift.

Marketing also uses refrain deliberately: a phrase in the headline, the subhead and the closing line. That works when the phrase is short and the spacing is wide, and it fails when it appears in every paragraph.

In both cases the rule is the same. Intentional repetition is exact and spaced. Unintentional repetition is approximate and adjacent. Recognising which you are looking at is most of the editing.

Everything else is a draft that has not been read by its author, and the fix is to read it: first sentences only, top to bottom, then a decision about each one.

What this covers

  • Outline extracted

    Paragraph openers read in sequence as an outline and merged where they repeat.

  • Openers varied

    No two consecutive paragraphs begin the same way, and generic connectives are gone.

  • One example per point

    No example carrying more than one claim, and no claim leaning on the same story twice.

  • Site-level check

    No two pages in a set could be swapped without a reader noticing.

  • Deliberate repeats kept exact

    Framing phrases and legal wording preserved verbatim, not paraphrased.

Questions this page usually raises

Does repetition make AI content detectable?
It is one of the more measurable signals. Openers and restated theses can be counted, but the stronger point is that repetition costs you readers regardless of how the text was produced.
Why does a model restate the same point?
It has no memory of the plan, so each sentence is generated from the text before it and a restatement is the safest next move. An editor's outline supplies the memory it lacks.
Is repeated content bad for search?
Within a page it wastes space and dilutes the topic. Across pages, near-identical drafts compete with each other and search engines typically pick one to rank.
How much repetition is acceptable?
Deliberate refrains are fine when they are short, exact and widely spaced. Accidental repetition is a finding at any density, because it always costs the reader something.
Should I delete the introduction or the conclusion?
Whichever restates the other. If the conclusion adds a decision or a next step, keep it and tighten the opening. If it only summarises, cut it.
Can the counter on this page catch conceptual repetition?
No. It counts the surface markers that usually accompany it: openers, self-references, summary loops and circular qualifiers. Conceptual repeats still need a human read of the outline.

Repetitive AI content

Run the checks, then decide who has to sign it.

The automated pass is free, needs no account and uploads nothing. Escalate to a named reviewer only where being wrong is expensive.