"Write me a 1,500-word blog post about email deliverability" produces something that reads fine and is worth nothing.
Not because the model is bad at writing. Because you asked it to generate a piece of content in a category it has already read ten thousand versions of, with no information it didn't already have — so what comes back is the statistical centre of everything ever written on the subject. Competent, generic, and interchangeable with the forty other articles doing the same thing.
The fix isn't a better prompt. It's a different division of labour: AI does assembly and structure, and a human supplies the things AI cannot have.
The short version
The workflow:
- Research — human-led, gathering real sources and your own material
- Brief — the highest-leverage document, and where quality is actually decided
- Draft — section by section, not all at once
- Fact-check — a dedicated pass, because this is where AI reliably fails
- Voice — making it sound like you rather than like everything
- Add what only you have — your data, your opinions, your examples
On the ranking question: Google's position, stated repeatedly since 2023 and unchanged, is that it focuses on the quality of content rather than how it's produced. There's no AI penalty. There is very much a scaled thin content penalty, and AI makes producing thin content at scale trivially easy — which is why the two get confused.
Why "write me a blog post" fails structurally
Worth understanding the mechanism, because it explains every fix below.
A language model produces the most probable continuation given its input. Ask for a blog post about a well-covered topic with no specific input, and the most probable output is a synthesis of the consensus — every point that appears in most articles on that subject, in the order they usually appear.
That's not a bug. It's the correct answer to the question you asked. You asked for the typical article. You got it.
Three consequences:
No information advantage. Everything in it existed already, which means an answer engine can produce the same summary without you, and a reader gets nothing they couldn't get elsewhere.
The characteristic shape. Every section the same length, every list three items, an introduction that restates the title, a conclusion that restates the introduction.
Confident errors. Where the model doesn't have a fact, it produces something plausible — and plausible wrong output looks identical to correct output.
The fix in one line: give it material it doesn't have, and constrain the shape.
Research stays human-led
The first and most important division.
What AI is genuinely useful for here: summarising sources you've selected, extracting the key claims from a long document, and organising material you've gathered.
What it shouldn't do: decide what's true, choose which sources matter, or supply the facts.
The reason is specific. When you ask a model for statistics or sources, you're asking it to recall rather than retrieve — and recall produces plausible-looking figures and citations that may not exist. Even with web search, it may select the first thing it finds rather than the most credible.
The practical workflow: you gather the sources. Ten browser tabs, three PDFs, your own notes, your own numbers. Then you hand those to the model and ask it to work with that material rather than its training.
And the highest-value input you have is your own: client results, your own audit data, the thing that surprised you last quarter. No model has that, which is precisely why it's valuable.
The brief decides everything
More consequential than the prompt, and the step people skip.
A brief is where you specify what this piece is, who it's for, what it argues, what it must contain, and what shape it takes. Without one, you're asking the model to make all those decisions — and its decisions default to the generic.
What a brief needs: the target reader and their specific situation, the question the piece answers, the position you're taking, the source material, the structure, the things that must appear, the things that must not, and the voice.
The eight inputs, with a worked example.
One thing worth saying plainly: writing a good brief takes real time — sometimes as long as writing a mediocre draft would. That's the trade. You're moving your effort from writing to specifying, and the return is that the specification is reusable while a draft isn't.
Draft section by section
Asking for a whole article produces a whole article's worth of averaged output.
Working section by section gives you control at each step, lets you feed different source material into different sections, and catches drift before it propagates through 1,500 words.
The practical method: brief the whole piece, get an outline, approve or fix the outline, then draft one section at a time with its own relevant sources.
It's slower per piece and produces something worth publishing, which is the only comparison that matters.
One habit worth building: after each section, ask yourself what a reader gets from it that they couldn't get from a summary. If the answer is nothing, that section is padding and should be cut or replaced with something only you can say.
The fact-check pass
Non-negotiable, and worth doing as a dedicated pass rather than while reading for flow.
Where AI reliably invents things: statistics, citations, quotes, dates, product features, and the specifics of how something works.
What makes it dangerous: the invented material is delivered with exactly the same confidence as the accurate material. There's no tell in the text.
The rule I'd hold to: every number, name, date, and claim about how something works gets verified against a source you can see. If you can't verify it, cut it — a piece with fewer specifics is better than one with a fabricated statistic, because one invented figure discovered by a reader costs more than the whole article earned.
The voice pass
The difference between "this is fine" and "this sounds like you."
Style prompts don't do much. "Write in a conversational tone" produces a model's idea of conversational, which is its own recognisable register.
What genuinely transfers voice is examples — your own past writing, given as reference material.
And some things you should just write yourself. Openings, transitions between major sections, and anything expressing an opinion. Those are where voice lives most visibly, and they're a small proportion of the total words.
What never to delegate
Three categories, and they're what make a piece worth reading.
Your own results and data. The numbers from your work. This is your only genuine information advantage, and it's the thing that makes a piece citable rather than replaceable.
Your opinions. Models are trained toward consensus and hedge instinctively. A clear position you can defend is something they structurally won't produce — and it's what readers remember.
Your examples. The specific client situation, the mistake you made, the thing that worked in your market and fails elsewhere. Situated, non-generalisable knowledge from having done the thing.
A useful test for any piece: what's in here that only I could have written? If nothing, you've produced something an AI could produce without you — which is precisely the content that has no future. Which formats survive.
Does AI content rank?
The evidence is clearer than the discourse.
Google's stated position hasn't changed since 2023: the focus is on the quality of content rather than how it's produced. There is no AI penalty in the ranking systems.
And the data supports it. Ahrefs analysed 600,000 pages and found 86.5% of top-ranking pages contained some level of AI-assisted content, with a correlation between AI content and ranking position of 0.011 — statistically indistinguishable from none. Purely human-written pages accounted for only about 13.5% of top-ranking content.
What does get penalised is scaled content abuse: producing content at volume with little value in order to manipulate rankings. That policy explicitly doesn't prohibit AI-generated content — it targets the behaviour regardless of how the content was made. AI is simply the cheapest way to do it, which is why AI-generated sites dominate penalty examples.
The distinction that actually predicts outcomes: not human versus AI, but whether the page offers something a summary can't replace.
Disclosure
Genuinely contested, so here's a defensible position rather than a rule.
Google doesn't require it. Its guidance is about content quality, not production method, and there's no ranking benefit or penalty attached to disclosing.
Some contexts do require it — certain publishing platforms, some client contracts, and some regulated sectors. Check your specific situation.
The test I'd apply: would you be comfortable if a reader knew exactly how this was made?
If yes, disclosure is optional and mostly irrelevant. If you'd rather they didn't know, that discomfort is telling you something about the work rather than about the disclosure policy.
And for a personal-brand site specifically: your name on a piece is a claim that you stand behind it. However it was drafted, you're accountable for every fact and every opinion in it. That's the standard that matters, and it's independent of what tools you used.
The pre-publish checklist
- Every statistic, name, date, and claim verified against a source I can see
- At least one thing in here that only I could have written
- A clear position, not a survey of all possible positions
- Opening and transitions written by me
- Reads like my other work, not like a model's register
- The first paragraph answers the question the piece is about
- Nothing in it that I couldn't defend if challenged
- Section lengths vary; not every list has three items
Frequently asked questions
Does Google penalize AI content?
No. Google's stated position, consistent since 2023, is that it focuses on the quality of content rather than how content is produced — there's no AI penalty in its ranking systems. What it does penalise is scaled content abuse: producing content at volume with little value in order to manipulate rankings. That policy targets the behaviour rather than the tool, and hand-written thin content is treated the same way. AI just makes producing thin content at scale far cheaper, which is why the two get conflated.
How do I make AI content sound human?
Style prompts barely work — asking for a "conversational tone" produces a model's version of conversational, which is its own recognisable register. What genuinely transfers voice is giving the model examples of your own past writing as reference material. Beyond that, write the openings, transitions, and anything expressing an opinion yourself, since that's where voice is most visible, and vary your structure deliberately because uniform section lengths and relentless three-item lists are the clearest tell.
Can AI content rank?
Yes, and most top-ranking content already involves it. Ahrefs' analysis of 600,000 pages found 86.5% of top-ranking pages contained some AI-assisted content, with a near-zero correlation between AI use and ranking position. What predicts performance isn't whether AI was involved — it's whether the page offers something a summary can't replace: original data, first-hand experience, a defensible position, or genuine depth.
Should I disclose AI use?
Google doesn't require it and attaches no ranking benefit or penalty either way. Some platforms, client contracts, and regulated sectors do require it, so check your specific situation. The useful test is whether you'd be comfortable if a reader knew exactly how the piece was made — if you'd rather they didn't, that discomfort is telling you something about the work rather than about disclosure. Either way, your name on a piece means you're accountable for every fact in it.
What to do next
Take a piece you're planning to write and ask the question before you open anything: what do I have that nobody else does?
A number from your own work. A client situation. A position you'd defend in an argument. An experience that contradicts the standard advice.
That's the piece. AI can help you structure and draft it. It can't supply it — and the pieces where you can't answer that question are the ones that were never going to be worth publishing.
Free: The content brief template.
Related guides
- The content brief that makes AI output usable — the highest-leverage step
- Fact-checking AI output — where it reliably fails
- How to make AI writing sound like you — the voice pass
- Does AI content rank? — the full evidence
- 6 content formats that still get clicks — what survives
- Human-in-the-loop — where review belongs
Join the Newsletter
Get practical marketing tactics delivered straight to your inbox.
Written by
Muhammad Basim
Related Articles
What Marketing Automation Actually Costs at Scale
Nobody's automation bill jumps because they added more automations. It jumps because they added steps to workflows that were already running — a filter here, an enrichment lookup there, a second notification — and the billing model charges for every one of them, on every run. That's the mechanic behind almost every "why did this […]
One Workflow, Three Tools: A Build Comparison
Feature tables don't tell you what a tool is like to use. The only thing that does is building the same thing twice and noticing where you got annoyed. So here's one realistic workflow — lead routing with enrichment and conditional assignment — specified once and mapped across Zapier, Make, and n8n. What changes between […]
GoHighLevel: 7 Automation Workflows to Build First
GoHighLevel gives you a hundred things you could automate, which is exactly why most accounts end up with forty half-built workflows and no measurable result. Seven are worth building first. They're the ones tied directly to revenue — catching leads before they cool, recovering appointments that would otherwise vanish, and asking for reviews at the […]