You have read AI slop. A post with a perfectly formed introduction that says nothing, five headings that could sit on any competitor's site, a paragraph of "in today's fast-paced world," and a made-up statistic near the end. It is grammatical. It is on topic. Nobody would put their name on it.
You have probably also read an AI draft that was fine. Specific, in the right voice, no invented facts, a couple of sentences a human tightened before it went out. Same kind of tool, very different result.
The tempting explanation is the model. It is almost never the model. Point two teams at the same model, and one of them ships slop while the other ships drafts, because slop is what a model produces when it is asked to write with nothing to write from. Three things separate the two outcomes, and none of them is a model setting.
Slop is what "write about X" produces
Ask any capable model to "write a blog post about inventory management for small retailers" and it will. It has no idea what you sell, who buys it, what your customers call things, or which of your competitors already owns that topic. So it writes the average of every inventory post it has ever seen. That average is slop by construction: it belongs to nobody because it was written for nobody.
The three failures readers notice most are all missing-input failures:
- Generic. The post could be on any site in the category. The model was not told what makes this business different.
- Off voice. Exclamation marks, emojis, "leverage," the reader addressed as "businesses." The model was not told how this company talks.
- Wrong facts. A price, a date, an integration, a customer quote, all plausible and all invented. The model was not told which facts it had, so it filled the gaps.
Every fix for slop is a fix for one of those three. The rest of this post takes them in the order they happen: what the model knows before it writes, the rules it writes under, and what a person does with the result.
1. Research: give the writer something to be specific about
A human freelancer who writes a good first post for you spent an afternoon reading your site, your pricing page and three competitors before they typed a word. A model needs the same afternoon, done once and written down.
In RankMill that afternoon is the research run. When you create a project, the research agent reads the website and the competitors you listed and writes a website profile: a plain markdown document with six fixed sections, Business, Products & services, Audience, Voice & tone, Positioning vs competitors, and Terminology & things to avoid. It runs once at creation and again only when you press "Re-run research"; nothing rebuilds it on a timer.
That document is what every later agent reads. The topic finder proposes topics from it. The writer gets it in full alongside the title and outline. The editor and the audit read it too. A post is specific because the profile is specific, which is why the profile is also the thing to fix when the posts are not. You can edit it by hand, or ask for an AI edit and review the change as a diff before it applies. The profile is never changed silently.
Facts the research cannot find
Research reads what is public. It cannot know the launch date you have not announced or the customer who agreed to be quoted. Slop fills those gaps; a draft leaves them empty or asks.
RankMill's agents are told, in the instructions every one of them receives, never to state a number, price, date, plan tier, customer name or product capability that is not in the material they were given. When a fact is missing they are told to drop the sentence that needed it, not to guess it and not to leave a placeholder. The topic finder goes one step further: it can attach up to four questions to a topic it suggests, asking for exactly the facts it would otherwise have to invent. Whatever you answer is handed to the writer as facts stated by the business. Whatever you leave unanswered is written around.
That last rule is the quiet difference between the two outcomes. Slop hedges a missing fact into a vague claim. A draft you can publish simply does not make the claim.
2. Rules: say how to write, one line at a time
The profile says what the business is. It does not say whether headings are sentence case, whether the reader is "you," or that em dashes are banned. Those are writing rules, and without them the model reverts to its default register, which is the register of slop.
In RankMill writing rules live on each project, one rule per line, set on the create form. Twelve suggestions are a click away, the generic habits most teams want: no em dashes, no AI cliches, active voice, short paragraphs, no fluff intros, no emojis, second person, concrete examples, sentence case headings, no exclamation marks, cite sources, and a practical takeaway. Each suggestion is stored as a full sentence, because "short paragraphs" leaves the model to guess what short means and "keep paragraphs to three sentences or fewer" does not.
Your own rules matter more than the suggestions. "Call it a workspace, never an account." "Never promise a delivery date." These are the lines that make a draft yours, and they are the ones only you can write. When a rule and the profile disagree, the rules win, and every agent that writes, edits or audits a post is told so.
Two things rules cannot do, stated plainly so nobody leans on them too hard. They are instructions in the prompt, not a linter: nothing rewrites the finished markdown to strip a banned word, so a slip is possible. And they cannot supply facts. A rule says how to talk, the profile and your answers say what is true.
3. The human pass: edit, do not approve
The third difference is the one teams skip when a deadline is close. A draft that was researched and rule-bound is close to publishable. It is not publishable. Someone has to read it, and reading is not the same as scrolling to the bottom and clicking approve.
A useful editing pass on an AI draft is short and has three questions:
- Is every fact one we can stand behind? Read each number, name and claim as if a customer will quote it back to you. If you cannot source it, cut it.
- Would we say it this way? Not "is it grammatical" but "is it us." Every sentence that makes you wince is either a missing rule or a missing line in the profile. Fix the sentence, then fix the source so the next draft does not repeat it.
- Is there a paragraph doing no work? AI drafts pad. Delete the paragraph and see whether anything is lost.
RankMill gives you two ways to make those changes. Press Edit on a finished post to open the markdown editor, change what you want and press Save. That is an in-place write: no agent runs, nothing is spent, the status does not move. Or tell the editor agent what to change in the conversation beside the article. It keeps its session across turns, so "make that shorter" on the third turn still knows what "that" was. Each turn costs one run and returns a list of the paragraphs it changed, with the old wording beside the new, an Undo on each, and a link that scrolls to the paragraph in the article. If a change could not be placed, the reply says so rather than pretending it landed. And a failed or stopped edit never turns a finished post into a failed one; the previous version stands.
The important thing is that the pass happens. The tool makes it cheap. It does not make it optional.
What the audit catches later
One post can be read carefully. Forty posts drift. The third post said the product integrates with Slack, the twentieth says it does not, and nobody noticed because nobody reads forty posts in a sitting.
The project audit is the backstop for that. It reads every finished post in a project, reports up to five of the most important problems, including contradictions between posts, drift from the profile and breaks from the writing rules, and then works through the fixes one task at a time. The audit costs one run and the fixes it makes afterwards are free. It is not a substitute for the human pass on each post, but it is the pass no human would do across the whole archive every month.
The short version
- Slop is what a model writes with nothing to write from. The fix is input, not a different model
- Research once, in writing: a profile of the business every draft is written from
- Missing facts get left out or asked for, never guessed
- Rules say how to write, one checkable line each, and they win over the profile
- Rules are instructions, not a linter, so read the draft
- The human pass checks facts, voice and padding, and feeds each fix back into the profile or the rules
- The audit catches what drifts across the archive
"It's not the model" is not a slogan. It is a checklist. A team that does the research, writes the rules and keeps the editing pass will ship drafts. A team that skips them will ship slop, whatever model they paid for.