It's not just X, it's Y: the sentence pattern on half the sites we scanned

There is a sentence shape almost everyone recognises and almost nobody has a name for. It denies a small claim so it can make a bigger one: it's not just a tool, it's a platform. The name for it is negative parallelism, though marketing teams tend to call it a hook.

Of the 80 sites whose stored page text we rescored under the current ruleset, 40 contained it at least once. Half. That puts it seventh among the 16 structural patterns SiteTell measures, ahead of uniform sentence length and far ahead of the generic openers people usually think of first.

The odd thing is that our own published figures used to say it barely existed. The reason why is worth a paragraph, because it is a lesson about measuring anything.

Why did this pattern look rare until recently?

Because it was not being checked. The rule was written, tested and fully working, but it lived only in a local seed file with no route to the production database. For a month every real scan ran without it, and our published dataset of 23,000 flags contained none of it.

On 2026-09-16 it reached production. Rescoring the stored page text from each site's latest scan, with the rule in place, gives the 50% figure above. That is the whole reason this page cites rescored numbers: a count of flags as they were originally stored would describe our deployment history, not anybody's writing.

What exactly does the rule catch?

Four shapes, which between them cover the common ways the figure gets written.

The first has a pronoun subject, as in , with or without "just". A bare pronoun is enough here, since hardly anyone writes that construction as a plain statement of fact.

The second allows any subject, as in , but insists on an elevation word: just, only, merely or simply. Without that requirement, an ordinary correction such as "the meeting moved to Monday" rephrased as a denial would be flagged, and at high severity, which would be expensive and wrong.

The third drops the negation and uses a comparison: . Because it reads as a compliment, it slips past editors more easily than the other three. A guard stops it matching quantities, so "we scanned more than 400 pages" never triggers it.

The fourth splits the figure across a full stop: Writers often reach for this version believing a full stop makes it less formulaic. It does not, and the rule treats the two sentences as one instance.

Between the halves, the rule accepts a comma, semicolon, colon, dash or hyphen. Curly apostrophes and typographic dashes are normalised first, since a content management system turns most straight quotes curly, and an earlier version missed live instances for exactly that reason.

Why does it read as AI-generated?

The research notes our ruleset started from describe it as one of the strongest and most specific tells, and the reasoning is easy to follow. The figure performs insight without supplying any. Denying the small claim implies the reader believed it, then the replacement arrives sounding like a revelation, while usually being vaguer than the thing it replaced. "A new way to run your clinic" says less than "a scheduling tool".

Models reach for it because it reliably sounds persuasive, and a great deal of copy now inherits it from them. When many pages on a site use it, readers stop hearing each instance as emphasis and start hearing a template.

Where does it show up?

In long-form writing more than anywhere else. Blog posts carry 0.374 instances per 1,000 words in the rescored text, against 0.082 on homepages, so the rate on blog posts runs about 4.5 times the homepage rate. The blog bucket spans 21 sites, and its heaviest single site accounts for 42% of the hits, which is high enough that the exact multiple should be read loosely.

That is the reverse of most vocabulary tells, which crowd onto product and homepage copy. The figure needs room to set up and pay off, and a blog introduction gives it both.

It also travels with a recognisable set of other habits. Pages carrying it are about 6.6 times likelier than an average page to contain a buried verb, and about 3.4 times likelier to open a sentence with a stock preamble, patterns spanning seven and ten sites respectively.

What does it look like, and what fixes it?

The examples below were written by us for this page, since quoting a real site could identify it. On each build of this site, every flagged version must trip this rule, and every rewrite must pass every rule SiteTell can check on a single passage.

The fix in each case is the one our rewrite engine is given: state the positive claim, once, directly. Then check whether the positive claim says anything. Often, as here, it turns out to be empty, and the real improvement comes from replacing it with a fact.

The comparison shape is the hardest to spot in your own writing, because it reads like modesty. Ask what the second half commits you to. If nothing, cut it and say what you do.

When is it fine to use?

Sometimes the denial is real information. If readers commonly believe something false about your product, correcting it is useful, and a reader will not mind the shape.

The rule knows this for three of its four shapes, which require either a pronoun subject or an elevation word. The split shape has no such guard. A plain correction like will be flagged, even though nobody would call it a flourish. We have left it that way deliberately for now, since the split version of the figure is common enough on real pages that loosening it would lose far more true hits than it saves false ones, but it is a known limit, and every flag quotes its sentence so you can dismiss those in a click.

Once per page is rarely what a reader notices. It is the third or fourth use that turns a device into a tic, and the scoring reflects that. The first three instances on a page each count in full at high severity, the top weight in the ruleset; after the third, further repeats are damped.

How to rewrite it

1. Find the positive half of the sentence and delete the negative half.
2. Read what is left. If it is a claim a reader could check, you are done.
3. If it is vague, replace it with the specific thing that makes it true: a number, a named feature, a customer's result.
4. If there was a genuine misconception worth correcting, state the correction plainly and give the evidence in the next sentence.