Ahrefs argues that slop is not a function of how much AI you use. We tested that on ourselves: twelve articles published between 27 August and 10 September 2026, nine produced by an AI pipeline and three written by hand, all scored by the same detector on the same text.
Table of contents
Key takeaways
- Across 12 articles published 27 August to 10 September 2026, the AI detector median was 7.3% for the nine pipeline articles and 5.7% for the three written by hand. The medians are close enough to be uninteresting.
- The spread is not. Pipeline articles ranged from 0.0% to 25.9%; hand-written ones sat between 3.6% and 6.6%. The machine is not worse on average, it is less predictable.
- The same text scored differently depending on what we fed the detector. One article read 17.6% as raw markdown and 4.5% as prose; another moved the other way, 17.7% to 25.9%.
- Our pipeline runs cost a mean of $0.83 and a median of 14.7 minutes, producing a median of six candidate drafts per article. Writing one by hand took over an hour of a person’s time and no model spend.
- No detector reading here is stable enough to gate a publishing decision on its own. We have been treating one as if it were.
What this is about
Ahrefs published its editorial process for AI-assisted writing, arguing that the volume of AI involved is not what makes content slop. We had twelve of our own articles and the metrics from producing them, so we checked what a detector actually sees when it compares machine-drafted work against hand-written work on the same site.
The design
Twelve articles published on this site between 27 August and 10 September 2026. Nine were produced by our generation pipeline: a donor text goes in, a persona and section prompt shape the draft, judges and repair passes run, and one of several candidate versions is chosen. Three were written by a person from the same donors and the same section rules, with no generation step at all.
Every article was scored by the same detector, ZeroGPT, on 11 September 2026. To make the comparison fair we stripped each article down to prose before scoring: no front matter, no table of contents, no headings, no tables, no link syntax. That matters, and the reason is in the numbers below.
There is no control group in the strict sense. Both groups were written to the same section prompts by the same editorial standard, but the topics differ, and nobody randomised which donor went to which method. This is a before-and-after comparison across a small sample, not an experiment. Twelve articles cannot separate method from topic.
Production cost came from the pipeline’s own run reports: wall-clock seconds, model spend in dollars, and the number of candidate versions generated per article.
The numbers
| Group | n | Detector median | Detector range | Words, median |
|---|---|---|---|---|
| Pipeline-written | 9 | 7.3% | 0.0 – 25.9% | 1,512 |
| Hand-written | 3 | 5.7% | 3.6 – 6.6% | 1,605 |
All twelve articles, prose only, scored 11 September 2026 on ZeroGPT. Lower is less machine-like by that tool’s scale.
The medians differ by 1.6 points on a sample of nine against three. That is not a finding. The spread is: the pipeline produced both the cleanest article on the site, at 0.0%, and the two worst, at 20.0% and 25.9%. Hand-written work clustered inside a three-point band.
The second table is the one that changed how we work.
| Article | Raw markdown | Prose only | Difference |
|---|---|---|---|
| Content recency study | 17.6% | 4.5% | −13.1 |
| Reddit and forums | 11.9% | 6.2% | −5.7 |
| GSC generative report | 14.2% | 7.7% | −6.5 |
| Web Almanac piece | 17.7% | 25.9% | +8.2 |
| 2026 priorities | 31.1% | 20.0% | −11.1 |
Same articles, two inputs: the markdown file as the pipeline scored it at generation time, and the prose extracted from it in September. Dates of measurement differ, which is part of the point.
Feeding the detector a table of contents, headings and tables moves the reading by up to thirteen points, and not always in the same direction. A tool that reports a single number to one decimal place is, on our texts, accurate to something like ten points.
Uniqueness behaved the same way. Across the nine pipeline articles the service returned readings from 54.03% to 98.68%, and re-checking an unchanged article days later produced a different number twice.
Production economics, nine pipeline runs: mean model spend $0.83, median wall-clock 14.7 minutes, median six candidate drafts per article. The three hand-written pieces cost nothing in model spend and took over an hour each of a person’s time, most of it on the measurement that goes into them rather than on sentences.
What could still be wrong
The sample is tiny and unbalanced. Nine against three, on a site two weeks old. A single outlier moves the pipeline median by more than the gap we are discussing. Treat every number here as an observation about twelve specific articles, not a rate.
Topics are not matched. The hand-written three all carry an original measurement we ran ourselves: robots.txt across 21 sites, JSON-LD across 20, our own titles against a best-practice list. Original measurement produces specific sentences, and specific sentences may be what the detector rewards. The method and the subject matter are confounded, and this design cannot separate them.
The detector is one tool, on one day. ZeroGPT is what our thresholds were built around, so we keep using it, but the raw-versus-prose gap above shows how much its output depends on input formatting. We have no second detector to triangulate against, and the vendor documents neither the model nor the scale.
Dates of measurement differ. The raw-markdown readings were taken at generation time between 27 August and 4 September; the prose readings all on 11 September. Detectors are updated without announcement. Some of the difference in the second table may be the tool changing, not the text.
And one belief of ours did not survive. The folk rule says em dashes are a machine tell and cutting them makes text read as more human. On the content recency article our pipeline logged the opposite: trimming dashes pushed the detector reading up, and the best score came from the version that kept fourteen of them. We have seen that pattern more than once. We have not tested it properly, so we report it as an observation, not a rule — but it is enough to stop us from chasing that particular tell.
What to do on Monday
Stop gating on a single detector reading. If one number moves thirteen points depending on whether you include your own table of contents, it cannot be the thing that decides whether a piece ships. Measure twice on identical input, and if the two readings disagree by more than a few points, believe neither.
Fix the input before arguing about the output. Score prose, not files. Strip navigation, tables, code and link syntax first. We were reading inflated numbers for two weeks and repairing text that did not need it, which is the same mistake as reading a metric without knowing what it counts.
Spend the saved time upstream, not on more drafts. Our median run generated six candidate versions for $0.83. Six versions of an article with nothing new in it is still an article with nothing new in it. The three pieces that scored most consistently were the three where a person decided what to measure first — and the threshold for measuring anything properly is higher than most sites clear.
FAQ
Does AI-written content score worse on detectors than human-written content?
On our twelve articles, not by median: 7.3% for nine pipeline pieces against 5.7% for three written by hand, all scored as prose on 11 September 2026. The difference between the groups was the spread, not the centre. The pipeline produced the single cleanest article and the two highest readings on the site. With nine against three, and with topics unmatched, this cannot be generalised beyond our own desk.
Why did the same article score differently on two occasions?
Two reasons we can separate and one we cannot. Input formatting: scoring the markdown file rather than the prose moved readings by up to thirteen points in our set, in both directions. Measurement date: the readings were taken two weeks apart, and detectors change without notice. What we cannot separate is how much of the remaining difference is simply the tool’s own variance on the same input, because we have not run that test.
Should we cut em dashes to sound less like a machine?
We stopped. On one article our own pipeline recorded the reverse effect: trimming dashes raised the detector score, and the best reading came from the version that kept fourteen. That is one observation, not a study, and we have not designed a proper test. But it is a reminder that stylistic tells are folklore until someone measures them, and that the tells are cheap to fake in either direction.
