Check an article against the fifteen rules we hold our own writing to.
Paste a draft or give us a URL. You get the score, every rule as its own row with the reason it passed or failed, and — the part no other checker does — the sentences that read like sourced facts with no source, quoted back at you.
A pasted draft is scored in your browser and never sent to us, so there is nothing for us to keep. A URL is fetched once as RankliBot, at most 10 times a day to any one site; neither the page nor the report is stored, and the only thing written down is a row saying a fetch happened. What RankliBot does · Privacy · Analyse a whole site
The rules
What each one checks, and why it is worth failing a draft over
These are not a checklist we wrote for a landing page. They are the gate in the product: an article that fails a hard rule never reaches a customer’s review queue, it goes back to be rewritten. Advisory rules lower the score and are shown, but they do not hold anything back.
Length. The body has to land between 1,500 and 3,000 words. Under the floor there is not enough of an answer to displace the pages already ranking; over the ceiling it is two articles sharing one title, and neither of them is finished.
Heading structure. Exactly one H1, at least two H2 sections, and no level skipped on the way down. The heading tree is how a crawler and a screen reader both work out the shape of a page, and an H4 sitting directly under an H2 tells both of them something that is not true.
Keyword in the title. The title contains the phrase the article is targeting. It is the line a searcher reads before they decide to click and the strongest single signal on the page; an article whose title omits its own target is aimed at nothing in particular.
Keyword in the opening. The phrase appears in the first hundred words of prose, before the first section heading. A reader decides in the opening whether this page is about their question, and a crawler weighs the same words heavily for the same reason.
Keyword in a section heading. At least one H2 contains the phrase. A section that names the subject is what lets a page answer the query in its own structure rather than only in its title, and it is where a featured snippet gets lifted from.
Keyword density. The phrase takes up no more than 2.5% of the article's words, counting the whole phrase each time it appears. Past that it reads as stuffing to a person before it does to a ranker, and the fix — a pronoun, a synonym, a shorter sentence — costs the writing nothing.
Banned phrases. The draft avoids every phrase on the brand’s own banned list. Those lists are where a company keeps the words it has decided never to say — a competitor’s name, a claim legal struck out, the marketing verb the founder hates — and a generator that ignores one publishes it under their byline.
Unsourced statistics, citations and quotes. No sentence pattern-matches an invented fact: a percentage next to attribution language, an academic-style citation with a year, a reference to “a 2023 study”, or a quotation attributed to a person. A language model asked for two thousand words will write “according to a 2023 study, 67% of teams…” with total confidence and no source, and that sentence goes out under your name — we cannot verify a claim, so we flag the shape of one and stop the draft until a person has looked.
Internal links. The article carries 2 to 5 contextual links to other pages on the same site. Internal links are the one ranking factor entirely inside your control, and the ceiling matters as much as the floor: a paragraph wearing five links is a link farm, not further reading.
Mentions what you sell. Somewhere outside the target phrase itself, the article mentions the category the business works in or one of its products by name. An article stamped out of a keyword can satisfy every structural rule while being about nothing — this is the check that asks whether it is about the business, and the keyword’s own occurrences are removed before it looks, so repeating the phrase cannot satisfy it.
Every section says something. Every H2 section, counting everything under its subheadings, carries at least 25 words. A heading that was planned and never written is the defect a reader spots instantly and a scorer that only measures totals cannot see at all.
The title’s promise. A title promising pricing contains a figure a reader could act on; a title promising a comparison names at least two things being compared. Advisory rather than blocking on purpose — “pricing” can honestly be written about pricing models, and a hard rule that misfires costs a real article — but a piece that promises numbers and delivers adjectives is the single most common way an AI draft wastes a reader’s time.
Meta description length. The meta description runs between 130 and 160 characters. Shorter leaves the snippet half empty, longer gets cut mid-word, and the snippet is the only sentence of yours most searchers ever read.
Title length. The title is 60 characters or fewer. Past that a search result truncates it, and the half that gets cut is the end — which is where a well-written title puts the thing that makes it worth clicking.
Readability. The Flesch–Kincaid grade level of the prose is 12 or below. It is a rough instrument, but it reliably catches the one failure that matters here: long noun-pile sentences that read like a press release rather than like somebody explaining something.
What this page cannot see
Three rules need something a stranger’s draft does not carry
They still run, and the report still shows them — with a line saying what was missing. A rule scored against an input we invented would be worse than no rule at all.
Your banned phrases
banned-phrases checks a draft against the words your brand has decided never to say. We have no list for you here, so nothing can fail it — the row says so rather than reading as a pass. In the product the list comes off your own site and stays editable.
What you sell
subject-coverage asks whether the article ever mentions your category or a product by name, outside the target phrase. It learns those from your site profile, so here it has nothing to compare against and drops to advisory — which is the rule’s own behaviour, not an exception made for this page.
The rest of your site
internal-links wants links to your other pages. A pasted draft names none of them and a URL check knows exactly one — the page it read — so links are still counted and classified, but the rule advises rather than blocks.
Straight answers
Questions people actually ask
Is this an AI content checker?
Not in the sense of guessing whether a machine wrote it. Nothing does that reliably, and tools that claim to are wrong often enough to get human writers accused. This checks the things AI drafts actually get wrong and that are measurable: invented statistics, headings with nothing under them, a target phrase repeated past the point of readability, an article that never once mentions the business it was written for.
If you want a number for “how much of this reads like a machine”, this is not that. If you want to know whether a draft is publishable, it is exactly that.
What does the fabricated-statistics checker look for?
Four sentence shapes: a percentage in the same sentence as attribution language (“according to”, “a study found”), an academic-style citation with a year, a reference to “a 2023 study”, and a quotation attributed to a person. Each flagged sentence is shown in full, so you can decide.
We cannot check whether a claim is true. What we can check is whether it is shaped like a claim that needs a source — and that catches the single most damaging thing a generator does, which is to write “according to a 2023 study, 67% of teams…” with total confidence and no study. A bare number or a year is deliberately not flagged: “since 2020” and “takes 15 minutes” are ordinary prose.
Does it work as an on-page SEO checker for a page I have published?
Yes — paste the URL. We fetch that one page, read the article out of its <article> or <main> element, drop the navigation, header and footer, and take the title and meta description off the page itself. The report says which element it read, because a page with neither gets read out of the body and the link counts may then include a sidebar.
It does not execute JavaScript. A page that draws its words in the browser sends us an empty shell, and we say so rather than reporting an article of nothing.
Do you keep my draft?
A pasted draft never leaves your browser: the scorer is a pure function, it ships in the page, and it runs in your tab. There is no upload to keep.
A URL check does reach our server, because your browser cannot fetch somebody else’s page for you. We fetch it, score it, return the report and keep nothing — the only thing written down is a row saying that a fetch happened, which is what stops this form being pointed at a stranger’s server all afternoon.
Is this really the same checker Rankli uses?
It is the same function, called the same way. When Rankli finishes writing an article it runs these fifteen rules on its own output and refuses to hand over a draft that fails a hard one. There is no “pro” version of the rules and nothing is held back here — a test in our build fails if a rule the scorer can produce is missing an explanation on this page.
What an account changes is not the checking. It is that something writes the article, checks it, and holds it back for you.
Why does it want 1,500 words?
Because that is the floor the product writes to, and the report has to agree with the product. It is a rule about long-form articles competing for a commercial query, not about every piece of writing — a 600-word release note is a good release note and a failing article by this measure, and the row will say so.
I left the keyword blank. What did you score?
The phrase your draft keeps coming back to: of every two-to-four-word phrase in it, the one that appears in more of your sentences than any other. We show it, label it as inferred, and mark all four rows it touched. We read it out of the prose and never out of the title on purpose: a keyword taken from the title would make “keyword in the title” pass by construction, and a check that can only pass is not a check.
Rankli runs these checks on everything it writes.
It reads your site, plans a month, writes 1,500-word articles in your voice — and then scores its own output against the fifteen rules above and holds the draft back when a hard one fails. You review what passed.