Content Writing with AI

AI can speed up writing many times over — but content written "at the push of a button" sounds generic, flat and soulless (what's known as AI-slop). The secret isn't to let AI write for you, but to write with you: it does the grunt work, you add the insight, voice and precision. Here's how.

The workflow — from idea to publish

1. Ideas: ask AI for 10 angles on a topic, and pick the best. You decide, not it.
2. Outline: build a skeleton — subheadings and key points. This is the step that determines quality.
3. Draft: let AI write to the outline, section by section. Several focused requests beat "write me an article."
4. Human editing: this is where the magic happens. Add real examples, opinion, personal experience. Cut filler.
5. Polish & SEO: title, meta description, keywords (see SEO with AI).
6. Repurpose: turn the article into posts, an email and a video (see Repurposing).

Tools — what to choose

• A general language model (ChatGPT, Claude, Gemini) — excellent for most needs, flexible and cheap. A good starting point for everyone.
• Dedicated writing tools (Jasper, Writesonic) — add templates, Brand Voice, SEO tools and team workflows. Worth it if you produce content at scale. See Jasper vs Writesonic.
• Prompts — quality depends on the prompt more than the tool. Use the prompt library and Prompt Engineering.

open_in_newTry Jasper open_in_newTry Writesonic

How to avoid AI-slop

• Give it a voice. Feed AI samples of your writing and tell it to mimic the tone.
• Add real experience. An example, a case, a number or an opinion only you can give. That's what AI won't know.
• Cut superlatives and filler. "In today's fast-paced world..." — delete. Get to the point.
• Vary your sentences. AI tends toward a uniform rhythm. Break it — short sentences alongside long ones.
• Verify facts. AI makes things up. Check numbers, quotes and links before publishing.

SEO & publishing

Good content needs to be found. Use AI to: research keywords, write titles and meta, and build a clear heading structure. But remember — Google rewards original, useful content, not volume. Quality over quantity. Details in the SEO with AI guide.

Why unedited AI writing all sounds the same

Worth understanding the mechanism, because it tells you exactly where to spend your effort.

A language model produces the most plausible continuation given everything it has read. On any topic that has been written about a lot, the most plausible continuation is the average of what everyone already said. That is not a flaw in the tool — it is what it is for. But it means an unedited draft is, structurally, a competent summary of the existing consensus.

Which explains the complaint about AI-written content perfectly. It is not badly written. It is indistinguishable — from other AI content, from the top ten results, from what a reader could already have found. Nobody dislikes it; nobody has any reason to read it.

So the useful reframing: your job is not to write better sentences than the model. It is to put things in the piece that are not in the average — your data, your mistakes, a customer's actual words, a test you ran, a position you will defend. Everything below is in service of that.

Sourcing is the work now

If the differentiator is material the model does not have, then gathering that material is the real job, and it happens before any drafting.

Where it comes from, in rough order of value:

• Something you did. A project, a test, a migration that went wrong. Specifics nobody can copy: what broke, what it cost, what you would do differently.
• Your own numbers. Not industry statistics — yours. "Our reply time went from two days to four hours" is worth more than any benchmark, and it is unarguable.
• Customer language. Support tickets, sales calls, reviews. How people actually describe the problem, which is never how the industry describes it.
• An interview. Twenty minutes with someone who knows the subject, transcribed, produces more usable material than an afternoon of research.
• A contrarian position you can actually support. Disagreeing with the consensus is the fastest way to write something that is not the consensus.

Then put that material into the prompt. Pasting your notes, your transcript and your numbers and asking the model to structure them produces something entirely different from asking it to write about the topic. It stops recalling and starts working with what you gave it, which is the mode it is genuinely good at.

The outline decides the article

The most consequential step, and the one people rush past to get to the draft.

Ask for an outline on a common topic and you get the generic skeleton: what it is, why it matters, types of it, benefits, how to start, common mistakes, conclusion. Every article on the subject has that shape, which is precisely why it is what the model predicts.

Three ways to break it:

• Lead with the thing you know that the generic version would put at position six, or omit. If your best material is the failure mode, the failure mode is section two.
• Make each heading a claim, not a category. "Pricing" is a bucket. "The pricing shape matters more than the price" is a section with something to say.
• Read the headings alone. If they tell a coherent argument on their own, the article works for the majority who skim. If they read as a table of contents, it does not.

Fix the outline before drafting anything. Rewriting a paragraph is easy; restructuring eight finished sections is a job people avoid, so a bad outline tends to survive to publication.

Draft section by section, not article by article

"Write the article" produces uniform output because the model is balancing the whole thing at once and averaging as it goes. Asking for one section, with its own brief, produces sharper writing and lets you steer between sections.

Per section, supply: the claim it makes, who it is for, roughly how long, what to include from your material, and what to leave out because another section covers it. That last one prevents the repetition that makes long AI drafts feel padded.

It also means you can change approach mid-article — one section as a worked example, one as a list, one as a straight argument — which is variation the model will not produce on its own and which makes a long piece readable.

Voice: what transfers and what does not

"Write in my voice" is the most requested and least effective instruction, because voice is not one thing.

What does transfer, if you supply it: samples — two or three pieces of your own writing pasted in, which works better than any description — and explicit rules: sentence length, whether you use contractions, first person or third, jokes or not, and a banned-word list. Specific constraints beat adjectives; "conversational" means nothing, "short sentences, no jargon, never say leverage" means something.

What does not transfer is the part of voice that is judgement — what you find interesting, what you refuse to say, when you go quiet. Those come out in the edit, which is another reason the edit is not optional.

If you produce content with a team, write the voice rules down once as a document rather than re-describing them each time. It is the same asset a dedicated writing tool sells you as Brand Voice, and a text file does most of the job.

The editing pass, in order

Editing generated prose is a different job from editing your own, because the failures are consistent. Doing it in this order avoids polishing sentences you are about to delete.

1. Cut the opening paragraph. Almost always context nobody needed. Check whether the piece is better starting at paragraph two — it usually is.
2. Cut the closing summary that restates what you just read.
3. Delete whole sections that say nothing. Generated drafts contain sections that exist because the outline had a slot. Be ruthless; length is not the goal.
4. Add your material — the example, the number, the story. This is the step that makes it yours, and the only irreplaceable one.
5. Break the rhythm. Models produce uniform sentence and paragraph lengths, and everything in threes. Merge some, split others, make one paragraph a single sentence.
6. Remove the vocabulary — unlock, elevate, seamless, delve, robust, game-changing, "in today's world", "it's important to note".
7. Cut the hedges. "Can potentially help to improve" has three softeners removing all the force. Say it or delete it.
8. Read it aloud. The last check, and the one that catches what the others missed.

Fact-checking, as a procedure

"Verify facts" is advice everyone gives and few operationalise. The failure is specific — fabricated material arrives in the same confident tone as everything true — so the check has to be mechanical rather than intuitive.

Go through the draft and mark every instance of four things: a number, a name, a date, a citation. Those are the categories that get invented. Then verify each against a source you found yourself, not one the model supplied — asking it for a source after the fact frequently produces a plausible reference to something that does not exist.

If a claim cannot be verified, you have two honest options: remove it, or state the uncertainty in the text. "We could not find a reliable figure for this" is a sentence real writers use and content mills never do, and readers notice.

One prompt-level defence worth using every time: tell it not to introduce statistics, studies or quotes unless you supplied them. It largely complies, and it turns fact-checking from a hunt into a short list.

SEO, honestly

Content has to be findable, and most SEO advice attached to AI writing is a decade out of date.

Keyword density is not a thing to optimise. What search engines say they reward — and what their behaviour is consistent with — is content that genuinely satisfies the query, from a source with some reason to be trusted, with evident first-hand experience where the topic calls for it. That last part is not a trick; it is why the sourcing section above is the most important one on this page.

What is still worth doing with AI: finding the queries people actually type, checking what currently ranks so you match the format readers expect, drafting titles and meta descriptions, and building a heading structure that answers the question directly. What is not worth doing: producing volume on the theory that more pages means more traffic. That strategy is what the recent ranking updates have been aimed at, and the sites that ran it are the ones that lost. More in the SEO guide.

Do you have to say it was AI-assisted?

A question that worries people more than it should, and the answer has two halves.

For search: the stated position of the major engines is that they care whether content is helpful and original, not how it was produced. Automation aimed at manipulating rankings is the thing they act against — not the use of a tool in the writing process.

For readers and clients: it depends on the expectation you set. A personal essay presented as your lived experience that a model wrote is a different thing from a documentation page drafted with assistance. If a reader would feel misled on finding out, say so. Many publications now have a stated policy, and if you write for someone else, ask what theirs is before you deliver — retroactively is a bad time to discover it.

The brief that produces a usable first draft

Most of the editing burden described above is created at the prompt. A brief with six things in it removes perhaps half of it, and it takes two minutes to write.

• The reader, specifically. Not "small business owners" but "someone who has tried this once, failed, and is sceptical". The second produces a piece that engages a real state of mind.
• The claim. One sentence stating what the reader should believe by the end. Without it, the draft covers the topic instead of arguing something.
• Your material, pasted in — notes, numbers, the transcript, the customer quotes.
• What to leave out. The obvious sections you do not want, the angles already covered elsewhere on your site. This does more work than any positive instruction.
• The constraints. Length, voice rules, banned vocabulary, and the standing instruction to invent no statistics, studies or quotes.
• The format of the output — one section at a time, with the heading as a claim.

Keep the brief as a template and fill it in per piece. After a few articles it becomes the most valuable file in your workflow, because it encodes everything you learned about what this particular model does wrong for your particular audience — and that is knowledge nobody else has.

When to write it yourself

• Anything short and high-stakes — a tagline, an apology, a difficult email. The register is the message and you will spend longer steering than writing.
• When the piece is your thinking. If you do not know what you believe yet, writing is how you find out, and outsourcing the draft skips the part that was the point.
• Anything where the specifics are the whole content — a post-mortem, a case study, a personal account. The model has none of them and will produce plausible filler between your facts.
• When you cannot check it. A subject you do not know well enough to spot a confident error is a subject to research, not to generate.

Knowing whether it worked

Volume makes people stop looking, which defeats the point of producing more. Three things worth watching per piece rather than in aggregate.

Time on the page and scroll depth tell you whether anyone read past the first screen — the honest measure of whether the opening earned attention. Returning readers tell you whether the writing is building anything. And replies, comments and shares tell you whether it said something; content that says nothing gets read and produces silence.

If a piece gets traffic and nothing else, it answered a query and left no impression. That is fine for a reference page and a failure for everything else — and it is the specific outcome that generated-and-lightly-edited content produces most reliably.

Common mistakes

• Publishing a raw draft. Unedited AI content = generic and suspect. Always edit.
• "Write me an article." A vague prompt = vague output. Give context, audience, tone and an outline.
• Ignoring facts. Publishing wrong information damages trust. Verify.
• Quantity over quality. 10 generic articles are worth less than one excellent one.