How to Use AI to Write Marketing Content That Sounds Human

Most advice about using AI to write marketing content stops at the prompt. You get a list of fifty prompts, you paste one in, and out comes something that could have been written for any company in any industry.

The prompt is maybe a fifth of the job.

The rest is everything that happens after the draft appears: the voice edit, the fact-check, and the first-hand detail only you can add. That part is where a useful piece of content is actually made, and it's the part almost nobody writes about.

Below is the workflow we use. It assumes you're a busy marketing manager or owner, not a prompt engineer, and it's built to be repeatable so the second article takes half the time of the first.

What AI is genuinely good at, and where it falls over

That last part matters more than anything else in this article. A model is optimised to produce the most probable sentence, and the most probable sentence is, by definition, the one everybody else has already written.

So use it for the jobs where "probable" is fine, and do the rest yourself.

Notice the pattern in that left-hand column. Every item is a transformation job, taking something that already exists and reshaping it. The right-hand column is all origination.

That's the line to hold. AI reshapes your material brilliantly and invents material badly, so your job becomes supplying the material and judging the output.

Average content isn't free, either. It costs you the hour you spent on it, the slot on your blog it occupies, and a small amount of reader trust every time somebody clicks through and finds nothing they couldn't have guessed. Published often enough, that adds up to a site people learn to skip.

We've written elsewhere about where AI actually helps small businesses at a strategic level. This piece is the task-level version of the same idea, applied to one job.

What you need before you write a single prompt

The single biggest quality difference between two people using the same AI tool is what they put in before they ask for anything. We call it a context pack, and you build it once.

• A voice sample of 500 to 1,000 wordsYour best-performing existing page, or something a founder wrote by hand.
• Five voice rules written as instructionsSentence length, first or third person, contractions, British or American spelling.
• A banned listThe phrases and patterns you never want to see in your copy.
• Real customer languageSales call notes, support tickets, review text. The words your buyers actually use.
• The source material for this specific pieceProduct docs, an interview transcript, your own bullet notes, real data you own.
• A one-line job for the articleWhat the reader should be able to do after reading it.

How to write voice rules that actually work

"Friendly but professional" is not a voice rule. It's a mood, and every model already thinks it's writing in one.

Useful rules are mechanical, because mechanical instructions are the only ones a model can follow reliably. Take your voice sample and measure it. How long is the average paragraph? Do you say "we" or "the company"? Do you ask questions or make statements? Do you use contractions?

Then write the answers down as commands. "One to two sentences per paragraph." "First person plural throughout." "Contractions in every paragraph." "British spelling." "State a position in the first line of each section, then give the exception."

Five rules of that kind will change an output more than a paragraph of adjectives.

Building the banned list is worth an hour of somebody's time. Read three AI drafts, highlight every phrase that made you wince, and write them down. Ours includes the openers that announce what the article is about to do and the closing paragraph that summarises what you just read.

How to write the prompt: a brief, not a wish

A weak prompt asks for an outcome. A strong prompt hands over a brief, the same way you'd brief a freelance writer you'd never worked with before.

The one instruction most people miss is the fourth band. Telling the model to use only the source notes you've supplied, and to flag anything it can't support, cuts the invented-fact problem dramatically. It doesn't eliminate it, which is why fact-checking still gets its own step later.

Here's the skeleton we use. Fill the brackets and keep it in a text file.

You are drafting a blog post for [business], a [what you do] for [who you serve].

READER: [who they are], who already knows [X] and is worried about [Y].
JOB: after reading, they should be able to [one specific outcome].
EVIDENCE: use ONLY the source notes below. If a claim is not supported
by the notes, write [CHECK] in square brackets instead of guessing.
VOICE: first person plural. Contractions. British spelling.
One to two sentences per paragraph. Take a position.
NEVER USE: [paste your banned list].
FORMAT: ~1,200 words. H2 subheadings. No introduction summarising
the article. No concluding summary.

SOURCE NOTES:
[paste transcript, product docs, customer emails, your own bullets]

Ask for the outline first, not the draft. Approve or fix the outline, then ask for the prose. Reviewing eight bullet points takes two minutes; rescuing 1,200 words built on the wrong structure takes an hour.

Three follow-up instructions do more than any amount of re-prompting from scratch.

The draft-then-edit workflow

The edit pass that removes the robot

Generic AI prose has a recognisable fingerprint. Once you can see it, you can't unsee it, and readers who spend all day online are starting to see it too.

1. Uniform paragraph lengthModel output tends towards three-sentence paragraphs, over and over. Break the rhythm. Put a five-word paragraph after a long one and watch the page come alive.
2. Lists of three everywhereFluent, forgettable and instantly recognisable. Cut one item or add a fourth, and make sure each item earns its place.
3. Hedged non-positionsSentences like 'there are many factors to consider' say nothing. Replace with the position you'd actually give a client, then name the exception.
4. No proper nouns and no numbersScan your draft for specifics. If a paragraph contains no name, figure, date or place, it's probably filler.
5. Restating the question before answering itThe model warms up before it says anything. Delete the warm-up sentence and start at the answer.
6. Words nobody says out loudIf you wouldn't use it on a call with a customer, take it out of the copy.

Here's what that looks like on a single sentence from a plumbing company's service page.

Same claim underneath. One version could belong to any trade business in the country, and the other could only belong to one that has thought about how customers get let down.

The test we use is simple. Could a competitor publish this paragraph with their logo on it and nobody would notice? If yes, that paragraph is doing no work for you.

Fact-checking: treat every claim as unsourced

This is the step people skip, and it's the one that can cost you actual money.

Language models generate text that looks like a citation with the same confidence they generate text that is one. A plausible statistic attached to a plausible research body, with a plausible year, is exactly the kind of output the technology is built to produce.

Build the check into the draft instead of bolting it on. Instructing the model to write [CHECK] wherever a claim isn't supported by your source notes gives you a search-and-fix list rather than a hunt.

Watch for the quieter version of the problem too. Invented statistics are easy to catch because they have numbers attached. Invented nuance is harder, and it shows up as a confident sentence about how something works in your industry that is subtly, plausibly wrong.

Anyone who knows the subject will spot it immediately, and those are exactly the readers you're trying to win.

For regulated topics, health, finance, legal, anything where being wrong harms somebody, we don't let AI draft the substance at all. It can structure, tighten and suggest headings. The claims come from a qualified human.

Is AI content bad for SEO?

Not inherently. Google's published position is that it rewards helpful, original content produced for people, regardless of how it was produced. What it targets is content generated at scale primarily to manipulate rankings, which its spam policies call scaled content abuse.

The practical translation is straightforward. Publishing forty thin AI articles a month to blanket a keyword set is the behaviour that gets penalised. Publishing four genuinely useful pages that happen to have had AI help with the first draft is not.

That first E is the whole problem with unedited AI output, and the whole opportunity if you edit properly. A model has no experience. It has never run your campaign, spoken to your customers or watched a launch go sideways.

Every specific you add is a signal the machine-generated version of your page can't produce. The client example. The number from your own dashboard. The thing you'd do differently now.

There's a second reason to care about specifics, and it's newer. AI-powered search results and assistants summarise pages rather than just linking to them, which changes what gets surfaced. We've unpacked how AI has changed search and what it means for the pages you publish.

Before you hit publish, run the usual on-page checks. Clear H2s, a description that matches the query, internal links, structured data where relevant. Our notes on on-page SEO for AI indexing cover what's changed and what hasn't.

Where this workflow breaks down

Three failure modes account for almost every disappointing AI draft we've seen, and none of them are solved by a better tool.

The first is thin input. If your source notes are four bullet points, the model fills the gap with generic knowledge, and you spend your editing time deleting rather than sharpening. When a draft comes back empty, the honest question is what you gave it.

The second is editing by regeneration. Rejecting a draft and asking for another one feels like progress because something new appears. You're sampling from the same distribution, so you get a different average article rather than a better one. Fix the brief or fix the text by hand.

The third is scaling before the process works. Getting one excellent piece out of this workflow tempts people into running ten at once, and quality collapses because the editing step is the bottleneck and nobody has ten spare hours.

Prove it on four pieces. Time yourself honestly, including the fact-check. Then decide what you can actually sustain.

How to tell whether it's working

Volume is the wrong measure. If AI has let you publish three times as much and nothing has moved, you've made three times as much noise.

That last one is the fastest signal you have, and it costs nothing. People who talk to your customers every day know instantly whether a page sounds like your company or like the internet.

If the numbers say the content is thin, the fix is almost never a better prompt. It's more of you in the draft.