What is AI slop? We read 13 rival blogs to draw the line

AI slop is low-effort AI-generated content pushed out in quantity with little regard for whether any of it is accurate, useful or wanted. Merriam-Webster made slop its word of the year for 2025 and defined it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence."

Notice what is missing. It says nothing about which tool produced the words. It is a claim about effort and volume, not authorship. That matters to us, because this post was produced with AI, the same as everything else on this blog.

So we read the blogs of 13 companies in our own corner of the market, hosting for the sites AI agents build, to find out which signs of slop survive contact with real pages. Several held up. One of ours did not, and we retract it below.

Where the word came from

The programmer Simon Willison argued for the word on 8 May 2024, crediting others with getting there first. His case: the internet already had a working model in spam, which never meant "sent by a machine" and always meant "sent to someone who did not ask." His test has the same shape, content that is "mindlessly generated and thrust upon someone who didn't ask for it."

The word is old and unkind. Soft mud in the 1700s, food waste in the 1800s, eventually anything of little value. By December 2025 it had been picked up widely enough that Merriam-Webster named it word of the year on the strength of how many people were looking it up.

The famous examples are absurd and visual: a fake photograph of a child clutching a puppy in a boat, circulated during Hurricane Helene; a band called The Velvet Sundown, streamed on Spotify before anyone established it was generative AI all the way down; a science fiction magazine that closed submissions because AI stories arrived faster than editors could read them. Those are the postcards. The version most people meet is quieter: a page that answers your search, reads fine, and leaves you with nothing.

Slop is not "made with AI". It is "published without a standard"

The tempting definition makes the tool the culprit: if a machine generated it, it is slop. That does not survive five minutes of use.

Wikipedia's article on the term notes that scholars argue no precise boundary exists. Charlie Guo's Field Guide to AI Slop, published in October 2025, opens with red herrings: academic vocabulary, flawless grammar, an absence of contractions, a fondness for the em dash. All things a careful human writer does, none of which tells you whether the page was worth publishing. Guo says plainly that he uses AI tools to draft and restructure his own work.

Detection tools do not rescue the definition either. Researchers at Northeastern University broke slop into seven measurable dimensions covering information utility, quality and style, and found roughly a third of their sampled text qualified. The finding that should end the argument is what happened when they asked large language models to apply the same guide: the models could not reliably identify slop, including their own output.

What is left is economics. Generating a page costs close to nothing. Checking one costs what it always did, a person's attention. Search engines and algorithmic feeds mostly reward the first number. So the working definition is not about tools at all:

Slop is what you get when nobody with a name decided that this particular page was worth publishing.

That version can be checked from outside, without knowing who typed what.

What we did, and what we cannot claim

Our research team profiled 13 companies in this niche on 22 June 2026, recording what was publicly visible: sitemap size, article counts, the dates in a sample of posts, whether anyone was named as author. Before publishing we opened those blogs again on 12 August 2026, because a seven-week-old snapshot of a live blog is exactly the stale artifact this post complains about. Several had changed a great deal.

What we cannot tell you: who or what wrote any of these pages. We have no access to drafts, prompts or editorial process, and whether a post is AI-generated is not observable from outside. Every observation below is about a published artifact, a date, a byline, a subject, and where we draw a conclusion from one, the conclusion is ours and arguable.

We read every competitor's blog so you don't have to

The first thing that falls apart is the assumption that everyone in a category is playing the same game.

here.now has no blog at all: 19 addresses in its June sitemap, with the effort going into machine surfaces instead, an llms.txt, a full-text version, an OpenAPI description, an agent.json. Static.app also has no blog and publishes more than twenty free tools. static.run has one article and an empty RSS feed. PageDrop has one article and about a hundred pages published by its users.

At the other end, Tiiny Host runs a blog sitemap of roughly 332 addresses once you count categories and pagination, and EdgeOne keeps around 40 resource articles inside a sitemap of 900 pages.

Between one article and three hundred, volume predicted nothing. What separated the useful blogs from the forgettable ones was visible in a single post.

The tells that held up

The subject comes from the news feed, not the product. On 12 August 2026 the blog index at dplooy, a service for publishing websites, showed 9 of 92 articles, and the recent ones were about Claude Opus 5, GPT-5.6, Kimi K3 and a model release from Anthropic. Our June snapshot found the same subjects on the same index, so this is standing policy rather than a slow week. It does not tell us how those posts were written. It tells us nobody appears to have asked which reader of a website-publishing service needed a parameter count.

A batch where a cadence should be. Publishing steadily is work; publishing twenty things at once is a button. Of the nine dplooy posts visible on 12 August, four carried the same date, 23 July 2026. A pattern worth noticing rather than a verdict, as the section below shows.

Nobody's name on it. Supadrop had 22 posts in June and about 31 by August, and on the day we checked only the featured post carried a byline: "Supadrop Team." Compare Stacktree, where posts are signed Steve Smith, Founder, and carry a visible "last updated" date. Anonymity is not slop on its own. What it removes is the person who would have to answer for the page being wrong.

Translation without an editor. OneClickLive mirrors its 23 English posts into Dutch. LiveSend publishes 14 guides in English and the same 14 in French and Spanish, none of them dated. Duplicating a library is a decision about coverage. Leaving every copy undated is a decision about how much anyone is expected to come back to it.

And here is what is not a tell. Not volume: Tiiny's large blog is one of the better ones in the set. Not clean grammar, not em dashes, not tidy formatting. Not comparison pages, which everyone here writes, including us. Those signals catch a register, not a standard.

What was not slop

We went in expecting a category full of filler and came out with a shorter list of complaints than we started with.

Stacktree went from around 19 posts in June to 43 in August, which by the batch logic above should look alarming. It does not, because the posts are narrow and awkward to fake: agents paying for their own hosting, a specific change to the protocol agents publish through and what breaks with it, what agents actually emit when they publish. You cannot write those without operating the product.

Supadrop, unsigned as it is, puts "honestly" in its comparison titles and then earns it, laying out competitors' pricing in plain numbers. Tiiny Host names a single author, Krupali Trivedi, on post after post across April to July 2026, the opposite of a faceless team byline. Seite builds its blog with its own static site generator, so the blog is the demonstration.

If a post about slop cannot find anything good in thirteen competitors, the post is the problem.

The observation we could not explain

Our June file on HTML Pub records 142 addresses in the sitemap, roughly 107 of them articles, and one line that mattered more than it should have: several posts share a single date, 27 May 2026.

Internally that sentence hardened into something it never said. "Several posts share a date" became "they publish in batches," and that was about to become a paragraph here.

So we checked. On 12 August 2026 the sitemap held 148 addresses, 121 under /blog/, and the modification dates were spread across months with no dominant day; 27 May accounted for two. A modification date is not a publication date, so that settles nothing on its own. The blog index is rendered in the browser and gave us nothing readable, so we opened four articles directly: dated 23 April, 6 May, 6 May and 12 May 2026, two of the four signed by named people, Michael Sacca and Yvonne Chow, the others under a team byline.

Four posts out of a hundred and twenty-one is a sample we would refuse from anyone else, and we are not extrapolating from it. What we can say is that we went looking for a batch dump and found named authors and spread dates. The claim does not survive, and we are withdrawing it.

Batch-dating is still a real signal. It is an observation about a publishing schedule, and the leap from there to "this was generated and dumped" is a hypothesis. We made that leap in private and caught it on the way to publishing.

How to spot it in thirty seconds

None of these detect AI. They detect the absence of a decision, which is what slop is made of.

1. Look at the dates on the posts either side of this one. A run of identical dates means the schedule is a queue, not a habit.
2. Look for a person's name. Then look for that name on other posts.
3. Ask whether the post answers a question anyone asked, or one a keyword tool produced.
4. Look for one fact the writer could only have got by doing something: a number they measured, a tool they used until it broke, a mistake they made.
5. Ask whether the subject belongs to the writer's own work or to this week's news.

A page can be entirely AI-written and pass all five. A page can be written by hand and fail every one.

Our own bar

We hold ourselves to four things, and this is the post where it makes sense to say them out loud.

Something real in every piece. A number we measured, a tool we used, an opinion we will defend, or a mistake we made. Our post on why pushing an index.html got hard is built on four quotes from people describing the problem in their own words. Our guide to connecting a domain is organised around the ways verification actually fails, because we collected them.

One name on the work. Every post here is signed Max Krou, and a person reads the final draft before it ships. AI drafts and restructures; the decision to publish is not automated.

An even cadence. The part we are less pleased about. Four of our own posts were staged to publish on a single day. We noticed while writing this section, which is a slightly humiliating way to find out, and spread them to roughly two a week. By the standard above, batch-dating our own launch would have been the exact tell this article describes.

No AI news. We will not write about model releases. There is real traffic in it, and it belongs to somebody else's readers.

None of this makes our posts good. It removes the ways a post can be empty without anyone noticing. If you want the argument this blog is for, it is that publishing a finished site should not be a project, and where we think a competitor is the better answer, we say so on the page.

Frequently asked questions