Slop is a judgement failure that a model happened to be present for

The page that arrives on a Friday

A client sends me a page and asks what I think. It is about their integrations. It is nine hundred words, correctly spelled, structured with headings, ending with a call to action.

It says nothing. Not incorrectly, nothing. Every paragraph is true and none of it would change a reader's decision about anything. The headings name topics. The examples are the kind of example that could appear on any company's page in any category. The closing line invites me to get started today.

The question I am being asked is whether it reads as AI written. That is not really the question worth answering, because a person could have written this and people have been writing it for twenty years. What went wrong is upstream of who or what typed it.

Three causes, none of which is a model

Nobody decided what the page was for. It exists because integrations should have a page. It has no specific reader, answers no specific question, and therefore has no way of being wrong, which is why it also has no way of being useful.

There was no source material. The person producing it did not have access to what a sales team gets asked about integrations, which integration questions block deals, or what the engineering team would say about how the integration actually works. With nothing specific to say, generic is the only available output.

There was no review that could reject it. Somebody read it and made small improvements. Nobody had the standing or the remit to say this page should not ship in this form.

Give any of those three to a model and the output improves substantially. Remove all three and a person produces the same page more slowly.

Where it shows up

It is worth being specific, because slop is usually discussed as a quality of prose and it is not confined to prose.

Copy. Sentences that assert a benefit without a mechanism. Our platform helps teams work more efficiently is not a claim, it is a shape where a claim should be.

Claims. This is the one that matters most and gets noticed least. A statement that sounds like a fact and cannot be sourced. Fast, secure, scalable, trusted by teams like yours. If somebody asked where that came from, is there an answer.

Layout. A page built from whichever blocks were available rather than from what the argument needed. Three feature cards because the component has three slots. An alternating image and text pattern continuing past the point where there was anything to alternate about.

Visuals. Illustrations that carry no information. A person at a laptop, an abstract shape, a screenshot cropped so nothing is legible. If removing the image loses nothing, it was never doing anything.

CTAs. Get started. Learn more. Ready to transform how you work. A CTA that would work on any page of any site is a CTA nobody wrote for this page.

Page structure. Sections that do not build. You can reorder most slop pages without loss, which is the clearest test available and takes ten seconds.

The reorder test

That last one deserves more than a line, because it is the fastest diagnostic I know and it needs no judgement about writing quality.

Take the page. Swap section three and section five. Read it again. If nothing is lost, the sections were not arguing, they were accumulating. A page that makes an argument has an order that matters, because each part depends on what came before it.

This works on pages a person wrote and pages a model wrote, and it works on your own writing, which is the uncomfortable part. It also points directly at the fix, which is that somebody has to decide what the page is trying to establish and in what sequence, and that is a thinking task rather than a writing task.

If you can reorder the sections and lose nothing, the page was accumulating rather than arguing.

Visuals are where nobody applies the test

Copy gets reviewed. Images almost never do, and the same three causes apply to them exactly.

The test is the same as for a paragraph: remove it and see whether anything is lost. A picture of a person at a laptop loses nothing. An abstract gradient shape loses nothing. A screenshot cropped so tightly that no text is legible loses nothing, and is worse than nothing because it implies there was something to see.

What passes the test is narrow and it is worth naming. A diagram of something that is genuinely hard to describe in a sentence. A screenshot where the specific thing being discussed is legible and pointed at. A comparison where seeing two states side by side is faster than reading about the difference.

The reason this matters commercially rather than aesthetically is that decorative images are the most expensive slop by area. They occupy the space where evidence could have been, on the pages where a buyer is deciding, and they push the actual information further down. A page with three stock illustrations has three places where it could have shown somebody something and chose not to.

The version of this that we hold ourselves to on this site is that every illustration is built from the same small set of shapes and each one says something specific about the article it sits with. That constraint is not about visual consistency. It is that a figure assembled from a vocabulary has to mean something, because there is no way to make a generic one.

Claims are the expensive category

Of the six, unsourced claims are the ones that cost real money, and they get the least attention because they read as normal marketing language.

The test is a question: if a buyer asked where this came from, is there an answer. Not a good answer, any answer. Enterprise grade security, where does that come from. Trusted by growing teams, how many and can we name any. Set up in minutes, has anybody timed it.

Run this on your own site and it is uncomfortable. Most sites contain a dozen sentences that would fail, and most of them were written years ago by somebody who no longer works there.

The failure mode is specific and it compounds. A buyer who catches one unsupportable claim discounts everything else on the page, including the parts that are true and specific and took real work. Which means unsourced claims do not just fail to help. They actively reduce the value of the accurate content around them.

Why generation makes this visible rather than causing it

I want to be precise here because the anti model version of this argument is both common and wrong.

Generation is very good at producing the shape of good writing. Structure, transitions, appropriate register, headings that look like headings. What it cannot supply is the specific thing your company knows that nobody else does, unless you give it to it.

So generation raises the floor on form and does nothing to the floor on substance. On a page with real source material and a real decision behind it, the result is often better than what a busy marketer would have produced. On a page with neither, you get a more polished version of nothing, faster and in greater quantity.

Which is why the useful response is not to write everything by hand. It is to fix the three causes, at which point generation stops being the variable that determines quality.

The review has to be able to reject

Most review processes cannot reject. Somebody reads the draft, suggests improvements, and it ships improved. That is editing, and editing a page that should not exist produces a better page that should not exist.

A review that works has one named person who can say this does not ship in this form, and whose saying so ends the discussion. Not a committee, not a process, a person with standing.

This is organisationally harder than it sounds, because rejection is socially expensive and the person doing it needs cover from somebody senior. In practice this is the single largest determinant of whether a site is good, and it has nothing to do with tooling.

The second requirement is that the reviewer sees the page against the site rather than alone. A page that is fine on its own and contradicts the pricing page is not fine, and that is not visible when reviewing a document.

A review pass that catches the six

This is the pass I run, and it works on pages whoever wrote them. It takes about fifteen minutes for a normal page and the first three items catch most of it.

The order matters. Purpose first, because if the page has no purpose the rest of the pass is polishing something that should be deleted.

Where teams get this wrong

Banning the tool. It addresses the fastest cause and none of the real ones, and it removes a genuine efficiency from teams that had the other two conditions in place.

Writing a style guide instead of assigning a person. Style guides describe how things should sound. They do not reject anything, and slop is rarely a style problem.

Reviewing the words and not the decision. The most useful review question is whether this page should exist, and it has to be asked before the draft, not after, because after the draft nobody wants to waste the work.

Measuring publication volume. If the team is measured on pages shipped, the review has an incentive against rejecting, and you have designed the process to produce the thing you are trying to prevent.

Assuming senior writers are immune. The three causes are structural. A very good writer with no source material and no decision produces a more elegant version of the same page.

What good looks like

A site without this problem does not read as a site with better writing. It reads as a site where somebody knew something and told you.

Concretely: pages state specifics that would be wrong if they were wrong. Numbers have sources. The illustrations show something you could not have inferred from the text. CTAs name what happens next in terms of this page's subject. And the site is smaller than you would expect, because things that had nothing to say did not get written.

The internal tell is that somebody can point at a page that was rejected in the last quarter. If nothing has ever been rejected, either everything is excellent or nothing is being reviewed, and the second is more likely.

The source material problem is the fixable one

Of the three causes, the second is the one most under your control and the one nobody works on.

Whoever writes the page usually cannot access what the company actually knows. The sales objections live in call notes nobody reads. The technical detail lives with an engineer who has no reason to talk to marketing. The reason a customer chose you lives in an onboarding conversation that was never written down.

Fixing that is not a writing project. It is building a route from the people who talk to buyers to the people who write pages, which is the same route the content gap work depends on. Once it exists, the pages get specific whether a person or a model drafts them, because there is finally something specific to say.

That is also the honest answer to the question I get most, which is how to make generated content sound less generic. You cannot, at the writing stage. The genericness was decided earlier.

Where our own work sits

Everything on this site is written by one of the two founders and rejected by the other when it is not good enough. That is the whole system and it is not scalable, which is fine at our size and would not be at forty people.

Creobot is aimed at the source material problem: collecting the questions visitors ask that no page answers, so the person writing a page has something specific in front of them. Creogen is aimed at the review problem at site scale, checking whether a page contradicts another page, which is the check a human reviewer cannot perform reliably across a hundred pages. Both are in private development.

Neither addresses the first cause, which is somebody deciding what a page is for. I do not think software addresses that one, and I would be sceptical of anything claiming to.

The argument, briefly

Slop is not a model output. It is what you get when nobody decided what a page was for, nobody supplied anything specific to say, and nobody could reject the result.

All three of those existed before generation and produced the same pages more slowly. What changed is throughput, which means a company with those three conditions now produces more of it, faster, and a company without them gets a genuine efficiency.

The work, then, is not about the tool. Decide what the page is for. Get the person writing it access to what the company knows. And give one person the standing to say no.