AI slop checker
Paste a draft. You get back the sentences that read as generated, the rule each one broke, and the fix. The rules are the ones that gate every article seodraft ships, run over your text alone.
What AI slop is
AI slop is text whose sentences are individually correct and collectively empty: a filler opener, a contrast that denies what nobody believed, a hedge on every claim the writer did not measure, a statistic with no owner, and a closing line that restates the paragraph. Most of that shape has a grammar, which is what lets code find it.
The rules came from a count. Over the eleven articles published from this repository, the contrastive negation alone appeared between 0 and 5 times per article, and eight of the eleven carried two or more.
How the checker works
Ten rules are deterministic: they read sentences, headings, paragraphs and lists, and return the same answer for the same text every time. Two more come from a classifier that reads meaning, and those two only advise. The same code runs inside seodraft as the editorial gate. Nothing on this page rewrites your text.
The twelve rules
Each rule has a name, a definition, the measurement behind it, and the fix as steps. Use the name when you brief an agent.
Contrastive negation
Blocks in the gate · Deterministic
Definition
A contrastive negation affirms one thing by denying another: "X, not Y", "it is not X, it is Y", "not just X but Y", or the same figure split over two sentences. The rule counts them across the prose, the title, the summary and the H2s, and allows one per article.
Measured
Measured over the eleven articles published from this repository, the shape appeared between 0 and 5 times per article, and eight of the eleven carried two or more.
Fix
Delete the negated half and read the sentence again. If the shorter sentence says the same thing, keep it shorter. Keep the contrast only where the reader believes the thing you are denying, and name that belief with its number.
Fails
The problem is not the prompt, it is the inputs. It is not about writing more, it is about writing what nobody else measured.
Passes
The problem is the inputs: the draft cites three industry figures and none of ours.
Filler phrases
Blocks in the gate · Deterministic
Definition
A filler phrase is a fixed expression that occupies a sentence without adding a fact: "in today's digital landscape", "it's important to note", "delve into", "game-changer", "en la era digital", "cabe destacar", "sin duda". The rule matches a bilingual list of 57 of them across the body, the title and the summary, and any hit fails.
Measured
Each entry has a spelling, so a string comparison finds it every time. The list carries 27 Spanish entries and 30 English ones, and a text in either language is read against both.
Fix
Delete the phrase. If the sentence still stands, you are done. If it collapses, delete the sentence.
Fails
In today's digital landscape, it's important to note that email deliverability is a game-changer for any startup.
Passes
Of the 1,240 emails we sent in March, 71 bounced.
Hedging
Advisory · Deterministic
Definition
Hedging softens a claim with "might", "perhaps", "somewhat", "arguably", "tends to be", "quizás", "probablemente" or "suele ser". The rule counts 15 such phrases over the prose and advises when they exceed 8 per 1,000 words.
Measured
The ceiling is a rate rather than a count, so a 2,000-word article can hedge sixteen times before it fires. Below that it stays quiet.
Fix
For each hedge, write down what you would need to know to remove it. If you know it, state it with the number. If you do not, say what you measured and stop.
Fails
This could potentially reduce bounce rates somewhat, and it might arguably be worth trying in some cases.
Passes
Cleaning the list cut the bounce rate from 5.7% to 0.9% on the next send.
Answer first
Blocks in the gate · Deterministic · Needs H2 headings
Definition
Answer first means the opening sentence under every H2 carries the answer to that heading on its own: at least six words, never a question, and never one of 43 openers that postpone the answer, such as "as we saw", "before we begin", "it depends" or "there are many ways".
Measured
The first sentence under a heading is the passage an answer engine extracts. The six-word floor is what keeps a one-word answer from passing.
Fix
Write the answer as one complete sentence. Use the rest of the section to qualify it. If the honest answer is a condition, put the condition and the outcome in the same sentence.
Fails
## How long does the trial last? Before we get into that, there are a few things worth understanding about how billing works.
Passes
## How long does the trial last? The trial lasts seven days from signup and ends without a charge if you cancel before day seven.
Intro length
Blocks in the gate · Deterministic · Needs H2 headings
Definition
The intro is every prose sentence before the first H2. The rule requires between 3 and 14 of them: a preview of the answer, one piece of proof, a preview of the structure and a transition.
Measured
Three is the smallest count that fits an answer, a proof and a transition. Across the twelve documents in this repository, intros run from 2 to 9 sentences.
Fix
Open with the answer in one sentence. Follow it with the number or case that proves it. Say what the sections cover. Start the first H2.
Fails
Email deliverability matters. ## What is deliverability
Passes
Seventy-one of the 1,240 emails we sent in March bounced, and every one came from a list we had not cleaned in a year. This article shows what we changed, in order, with the bounce rate after each step. Start with the list, because that is where 90% of the damage was. ## What is deliverability
Generic closing
Blocks in the gate · Deterministic · Needs H2 headings
Definition
A generic closing is a last H2 titled "Conclusion", "Final thoughts", "Summary", "Wrapping up", "Conclusiones" or "Reflexiones finales". The rule checks the last heading against 13 of these and fails on a match.
Measured
The rule reads the last H2 only, so a section with one of these titles in the middle of an article passes.
Fix
Name the decision the reader has to make now. Give the next step and one resource for it. If there is no decision and no next step, delete the section.
Fails
## Conclusion In summary, deliverability is a complex topic with many factors to consider.
Passes
## Clean the list before you touch the template Run the bounce report for the last 90 days and remove every address that bounced twice. Then send the next campaign and compare.
Sentence variety
Advisory · Deterministic
Definition
Sentence variety measures how much sentence length moves across a text. Over 15 or more sentences, the rule computes the standard deviation of words per sentence and advises when it falls under 4.
Measured
Across the twelve documents in this repository the deviation runs from 4.4 to 9.6 words. The threshold of 4 sits under the lowest of them.
Fix
Find three consecutive sentences of the same length. Cut one to under eight words. Let another run past twenty-five if it has two clauses that belong together.
Fails
We cleaned the list in March. We changed the sender domain in April. We rewrote the template in May. We measured the bounce rate in June.
Passes
We cleaned the list in March. April was the sender domain, which took two weeks of DNS changes and one support ticket nobody answered. The template came in May.
Paragraph length
Advisory · Deterministic
Definition
Paragraph length counts the sentences in each prose paragraph and advises when any paragraph runs past four. Lists, headings, tables and code are skipped.
Measured
Eight of the twelve documents in this repository keep every paragraph at four sentences or fewer. The advisory names the longest paragraph it found.
Fix
Find the sentence where the subject changes and break the paragraph there. If no sentence changes the subject, the extra sentences repeat the first one, so delete them.
Fails
The list had 1,240 addresses. We had not cleaned it in a year. Seventy-one bounced in March. The sender domain was also new. It had no reputation yet. Both problems hit the same campaign.
Passes
The list had 1,240 addresses and we had not cleaned it in a year. Seventy-one bounced in March. The sender domain was also new, with no reputation yet. Both problems hit the same campaign.
Bold-label lists
Advisory · Deterministic
Definition
A bold-label list is a bullet list where each item opens with a bold label and a colon: "**Speed**: the fastest option". The rule counts bullets in that shape and advises when they exceed 70% of a list with four or more items.
Measured
The shape is the default output of a model asked for key points. The 70% line lets a list keep one or two labelled items among plain ones.
Fix
Keep one list in that shape if the labels are real categories. Turn the others into prose, or into a table if every item has the same two or three attributes.
Fails
- **Speed**: the fastest option we tested - **Cost**: the cheapest per send - **Support**: answered in under an hour - **Docs**: complete and current
Passes
It was the fastest option we tested and the cheapest per send. Support answered in under an hour, and the docs covered every endpoint we called.
Thin content
Advisory · Deterministic · Needs H2 headings
Definition
Thin content is an article whose prose, with code blocks and headings removed, runs under 900 words. The rule counts the words and advises below that line.
Measured
The threshold is editorial and the rule advises only. The twelve documents in this repository run from 517 to 1,364 words, and two of them sit under 900.
Fix
Add the measurement, the case or the mistake that is missing. If you have nothing to add, the article is finished.
Fails
A 480-word article with three H2s and no numbers.
Passes
The same article with a table of the bounce rate after each of the four changes.
Straw-man contrast
Advisory · Classifier
Definition
A straw-man contrast is a contrastive negation whose negated half nobody held. A classifier reads the article and returns the probability that at least one contrast is of this kind. Above 0.6 the tool advises.
Measured
The question scored 0.90 on an article whose contrasts deny nobody, against 0.43, 0.50 and 0.55 on three articles from this repository whose contrasts name a real belief. The 0.6 threshold sits in that gap.
Fix
For each contrast, write down who believes the negated half and what number they would quote. If you cannot name them, delete the negated half. If you can, put their number in the sentence.
Fails
Deliverability is about reputation, not about luck.
Passes
Deliverability is about reputation. The four founders we talked to blamed the template first, and the template was the last of four things that moved the number.
Unsourced numbers
Advisory · Classifier
Definition
An unsourced number is a statistic with no owner: "studies show 73% of users", "most companies see a 3x lift". A classifier reads the article and returns the probability that its figures are first-hand measurements or named cases. Below 0.35 the tool advises.
Measured
Three numbers with no owner satisfy every rule that counts evidence. A draft engineered to pass the whole deterministic standard on invented statistics scored 0.04 on this question, where this repository's reference article scores 0.90.
Fix
For each figure, name who measured it and when, or replace it with a figure you measured. Delete a figure you can do neither with.
Fails
Studies show that 73% of marketing emails are never opened, and most companies see a 3x lift after cleaning their list.
Passes
Of the 1,240 emails we sent in March, 412 were opened. After cleaning the list, the April send reached 1,098 addresses and 468 opened.
What it does not flag
- A concessive contrast passes the checker with no finding. "Retrying did not recover the screen, but a reload did" is a fact from one of seodraft's own articles, and a rule that fails an article for reporting a fact gets worked around.
- An aphoristic closing line passes. A detector for it was measured over the 82 short paragraph closers in this repository and flagged 7, of which 3 were codas and 2 were already caught by the contrast rule. At 43% precision it was dropped. A classifier question was tried next and scored two plain facts at 0.85 and 0.91, so it was dropped too.
- Authorship is never judged. The checker names sentences a reader will recognise as generated, and a human writes them too.
For agents and scripts
POST the text and read the findings back as JSON, or hand an agent the Markdown twin of this page and the skill file, which carries the rules as instructions to follow before drafting.
POST /api/slop
curl -s https://seodraft.app/api/slop \
-H "Content-Type: application/json" \
-d '{"text":"# Title\n\nYour draft here."}'- This page as Markdown
- Skill file for agents
- JSON endpoint
Questions
Does it give an AI-detection score?
No. It returns the sentences that broke a named rule, with the rule and the fix.
Which languages does it read?
English and Spanish, in the same run. The filler list, the hedge list and the contrast grammar carry both.
Does this page pass its own rules?
A test in the seodraft repository runs the rules over the Markdown twin of this page. A definition may quote its own examples. Everything else on the page has to pass.
Is the text stored?
No. The rules run and the text is discarded. When the classifier layer is on, the text is sent to TypeSafe for two yes/no questions and is not retained there either.