Most advice on how to write SEO articles is either a list of ranking factors with no writing in it, or a writing lesson with no ranking in it. What follows is the process we encoded into a product: eight steps, then fifteen checks that decide whether the piece is publishable. The checks matter because they turn taste into something you can apply at nine in the morning on the fourteenth article of the month, when your judgement is worse than it was on the first.
How to write SEO articles in the order the work actually happens
The order is not decorative. Every step below removes a decision from the next one, and skipping one of them is how a draft ends up structurally fine and useless.
Step 1: pick a phrase you could plausibly win
Pick for intent before volume. A phrase with a buyer behind it and modest volume beats a big informational phrase you will never convert. Look at what already ranks: if page one is nine established brands and a Wikipedia entry, you are shopping in the wrong aisle this quarter.
Write the phrase down exactly as a person would type it. That exact string is what goes in the title, the opening and a heading later, and it is the thing everything else is measured against.
Step 2: read the first five results before you write anything
Twenty minutes. You are looking for three things: what the page that ranks is actually answering, what every one of them covers, and what none of them covers. The last one is your article. If you cannot find a gap, either you have not read carefully enough or the phrase is not worth your month.
Step 3: outline in headings, then check the outline for holes
Write the H2s first. Each one should be a question a reader would ask in that order, and each should be answerable in 150 to 300 words. If a section cannot carry 25 words of real prose, it is not a section, it is a sentence in another section.
Two H2s is the floor for a long piece. Six to nine is normal for a 2,000-word article. Add H3s only where a section genuinely splits, and never jump from an H2 to an H4, because that misrepresents the structure to a screen reader and a crawler alike.
Step 4: write the opening as the answer
The first 100 words decide whether a reader stays and tell a crawler what this page is. So answer the question there, in plain words, including the exact phrase you chose in step one. Then earn the rest of the article.
The opening you want is the one that would satisfy a reader who leaves immediately. Counter-intuitive, and it is the single biggest difference between pages that hold attention and pages that get a back-click at four seconds.
Step 5: write sections that stand alone
Nobody reads these top to bottom. They scan the headings, drop into the one that matches their question, and leave. So each section should make sense cold, restate just enough context to be readable on its own, and stop when it has answered its own heading.
Keep sentences short. Twenty words or fewer is a good target. Say the specific thing: a number, a name, a mechanism. "Fifteen checks on every draft" beats "rigorous quality control" every time, in prose and in a search result.
Step 6: place internal links while the paragraph is in front of you
Two to five links to your own article-like pages, put in the sentence where they are genuinely useful. Not your terms page, not your homepage, and not a block of "related reading" nobody clicks. If you cannot think of a relevant page, that gap is your next article.
Step 7: run the checks before you publish, not after
This is where taste becomes a checklist. The fifteen rules below are the ones Rankli runs on every draft, and eleven of them block publication.
Step 8: publish, then leave it alone
Six weeks minimum before you judge anything. Search results move slowly, and the urge to rewrite a page in week two is the most reliable way to learn nothing from it. Write down what you expect the page to do before you publish it, then check that prediction on a fixed date. A prediction made in advance is worth more than a chart read afterwards, because you cannot argue yourself into having expected whatever happened.
The eleven checks that block a draft
Word count between 1,500 and 3,000. Heading hierarchy: exactly one H1, first, no skipped levels. Keyword in the title, in the first 100 words, and in at least one H2. Keyword density at or below 2.5% of total words. Banned phrases: your own list of words you will not publish. No fabricated specifics: any sentence shaped like a sourced claim, a percentage next to attribution language, a citation with a year, a quotation attributed to somebody, fails. Internal links, two to five. Subject coverage: with every occurrence of the keyword removed, does the article still mention what you sell? Section substance: no H2 under 25 words of prose.
Eleven rules, and none of them is about elegance. Each one is a defect you can find without reading the article twice, which is exactly why a machine should be finding them.
The four that advise
Title length at or under 60 characters, or it gets truncated in results. Meta description between 130 and 160 characters. Reading grade at or below 12 on Flesch-Kincaid, which mostly catches sentences that have grown three clauses too long. Title delivers: if the title promises a price, there had better be a figure; if it promises a comparison, name at least two things.
These lower the score and never block, because each has a real exception. A technical audience tolerates grade 14 prose. A title can be 64 characters and still be the right title. A rule that misfires costs a piece of work, and a nit that shows up on a scorecard costs nothing.
What this looks like automated
Every post on this blog carries its own scorecard, including the one you are reading. The process above is exactly what Rankli runs for a customer: it researches the phrases, plans a month with mixed intents, writes each piece against the outline, places the internal links, then scores the finished draft against these fifteen rules. A hard failure regenerates the draft once and is written onto the article; switch on "hold drafts that fail checks" and nothing that fails ever leaves the app.
You can run the same fifteen checks on a draft you wrote yourself, free, in the SEO article checker. If you want the longer version of what each rule is protecting you from, we wrote that up in what to look for in generated output.
The five mistakes that cost the most time
Writing before reading the results. Twenty minutes of reading page one saves a rewrite. Skipping it is how you produce the tenth version of an article that already exists nine times.
Burying the answer. A 300-word warm-up before the point loses the impatient reader and gives a summariser nothing to lift. Answer first, elaborate second.
Chasing volume over intent. A phrase with 8,000 searches a month and no buyer behind it produces traffic you cannot use. Pick the phrase your customer types the week before they buy.
Padding to hit a word count. The 1,500-word floor exists because commercial phrases need substance, not because length is a virtue. If the article says everything it has to say in 1,600 words, that is the article. Padding shows up as thin sections and long, empty paragraphs, and both are visible to a reader before they are visible to a scorer.
Rewriting in week two. Nothing has settled. Change one thing at a time and give it six weeks, or you will never know which change did what. This is the same discipline that makes a control group worth having.
None of these is exotic. They are what happens when a person is producing their fourteenth piece of the month and their judgement has thinned out, which is exactly the condition a checklist exists for. Full plan detail, if you want the pipeline to run the checklist for you, is on the pricing page.
The three habits that matter more than any single article
Cadence beats brilliance. Thirty pieces a month covering a cluster properly will outperform four beautiful essays, because search rewards coverage and because you learn faster with more attempts on the board.
Coverage beats keywords. Own the whole question set around a buying decision, not the one phrase with the highest volume. The internal links between those pieces are worth as much as the pieces.
Measurement beats opinion. Read your own Google Search Console rather than a tool's estimate, and be willing to conclude that an article did nothing. And now that buyers ask assistants as well as search engines, measure that too: we track how ChatGPT, Perplexity, Gemini and Claude answer buying questions in our category, as an experiment with a held-out control rather than a chart, and the method is written up on the proof page.
Do those three for two quarters and the individual articles stop mattering very much, which is the point.
Rankli's $1 trial writes three of these for you, in your voice, published live to your own site, each with the full fifteen-check scorecard attached. Fourteen days, cancel from the app any day, unused articles refunded.